{ "schema_version": "1.0", "entry_count": 117, "coverage": { "entry_count": 117, "oldest_date": "2026-01-08", "newest_date": "2026-08-26", "checked_nightly_from": "2026-07-27" }, "entries": [ { "id": "E-2026-114", "date": "2026-08-26", "vendor": "Anthropic", "headline": "Sonnet 5 sessions now run 33,000 tokens longer before compacting", "what_plain": "Claude Code compacts a long session automatically once it approaches the model's context limit. For Sonnet 5 on the 1M-token window that point was about 934,000 tokens; version 2.1.247 moved it to about 967,000.", "quote": "Changed Sonnet 5's default auto-compact window to its full 1M context, so sessions on the 1M window now auto-compact at about 967K tokens instead of about 934K", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/manual_captures/2026-08-27/ant-claude-code-changelog-20260827.html", "bundle_sha256": "38144be62f6056c9fd30686dd3933b331ad8ab91d462d43cf0387db9f17a3060", "source": "reported", "date_note": "2026-08-26. Anthropic's own date: the changelog carries this version as ." }, { "id": "E-2026-124", "date": "2026-08-26", "vendor": "Anthropic", "headline": "Sign-in enforcement now exits when it cannot read managed settings", "what_plain": "When Claude Code could not read an administrator's managed settings, it fell back to host-supplied or per-user Windows registry settings and carried on. Version 2.1.247 made it exit at startup instead, so organisation sign-in enforcement cannot be skipped by an unreadable settings file.", "quote": "Changed organization sign-in enforcement to exit at start when the administrator's managed settings cannot be read, even if host-supplied or per-user Windows registry settings exist", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/manual_captures/2026-08-27/ant-claude-code-changelog-20260827.html", "bundle_sha256": "38144be62f6056c9fd30686dd3933b331ad8ab91d462d43cf0387db9f17a3060", "source": "reported", "date_note": "2026-08-26. Anthropic's own date: the changelog carries this version as ." }, { "id": "E-2026-125", "date": "2026-08-26", "vendor": "Anthropic", "headline": "Analytics now stay off from startup, not just after login", "what_plain": "Where managed settings force gateway login or a custom OAuth deployment is configured, Claude Code switched analytics off only once the user had logged in, leaving a window at startup where they ran. Version 2.1.247 keeps them off from startup.", "quote": "Changed analytics to stay off from startup, not only after login, when managed settings force gateway login or a custom OAuth deployment is configured", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/manual_captures/2026-08-27/ant-claude-code-changelog-20260827.html", "bundle_sha256": "38144be62f6056c9fd30686dd3933b331ad8ab91d462d43cf0387db9f17a3060", "source": "reported", "date_note": "2026-08-26. Anthropic's own date: the changelog carries this version as ." }, { "id": "E-2026-112", "date": "2026-08-25", "vendor": "Anthropic", "headline": "Computer use on macOS lost its Finder and Dock exemption", "what_plain": "Claude Code's computer-use feature on macOS could click the desktop, the Dock and Finder windows without the user granting access to Finder. From version 2.1.243 those clicks require the same permission dialog as any other application.", "quote": "Changed computer use on macOS so clicking the desktop, Dock, or a Finder window requires granting Finder via the access dialog, like any other app", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-26/ant-claude-code-changelog", "bundle_sha256": "6e03371d6037d060be81bd89fbf8f745a1d41ed5a316cbc5bd71f3a8e5bb67e3", "source": "reported" }, { "id": "E-2026-120", "date": "2026-08-25", "vendor": "Anthropic", "headline": "The sandbox prompt stopped telling Claude which hosts were allowed", "what_plain": "The sandboxed Bash tool's prompt used to carry the list of network hosts Claude was allowed to reach, so Claude treated any host not on the list as blocked and did not try. Version 2.1.243 removed the list from the prompt; Claude now attempts the request and the user can approve a new host.", "quote": "Changed the sandboxed Bash tool prompt to no longer list allowed network hosts, so Claude attempts requests (and you can approve new hosts) instead of assuming unlisted hosts are blocked", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-26/ant-claude-code-changelog", "bundle_sha256": "6e03371d6037d060be81bd89fbf8f745a1d41ed5a316cbc5bd71f3a8e5bb67e3", "source": "reported" }, { "id": "E-2026-121", "date": "2026-08-25", "vendor": "Anthropic", "headline": "Model and effort commands stopped waiting for the turn to end", "what_plain": "On Bedrock, Vertex AI and Foundry, and whenever telemetry was switched off, the /model, /fast and /effort commands were queued and only took effect once the current turn finished. Version 2.1.243 made them take effect immediately on those platforms too.", "quote": "Changed /model, /fast, and /effort to also run immediately instead of queueing until the turn ends on Bedrock, Vertex, and Foundry and when telemetry is disabled", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-26/ant-claude-code-changelog", "bundle_sha256": "6e03371d6037d060be81bd89fbf8f745a1d41ed5a316cbc5bd71f3a8e5bb67e3", "source": "reported" }, { "id": "E-2026-123", "date": "2026-08-25", "vendor": "Anthropic", "headline": "An idle session now starts at most three check-ins per goal", "what_plain": "Version 2.1.239 spaced out /goal's repeat check-ins but set no ceiling, so an idle session could keep starting them indefinitely. Version 2.1.246 capped it at three per goal, with the user's next message allowing three more.", "quote": "/goal: Changed idle sessions to start at most three check-ins on long-running background work per goal; your next message allows three more", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-26/ant-claude-code-changelog", "bundle_sha256": "6e03371d6037d060be81bd89fbf8f745a1d41ed5a316cbc5bd71f3a8e5bb67e3", "source": "reported" }, { "id": "E-2026-111", "date": "2026-08-22", "vendor": "Anthropic", "headline": "Every \"generally available\" claim was rewritten across two years", "what_plain": "Generally available is a term of art in software. It means a product has left testing and is fully supported for production use, the point at which a company takes on support obligations. Anthropic's API release notes used that phrase across two years of entries. Between 14 and 22 August 2026 every instance was replaced with wording that makes no such claim: out of beta, widely released, available, no longer requires a beta header. The dates on the entries were not changed, so the record still reads as though it were written at the time. The entries touched run from 30 May 2024 to 28 May 2026.", "quote": "our most capable generally available model", "artifact_url": "https://platform.claude.com/docs/en/release-notes/overview", "our_archive": "ledger/manual_captures/2026-08-27/anthropic-relnotes-BEFORE-20260814.md", "bundle_sha256": "4eebb61915c3db72b2735385190393b5194e43f60fa17562e14494a16dea67ea", "source": "reported", "date_note": "2026-08-22. Wayback brackets the edit rather than pinning it: the phrase is present in the 14 August capture and absent from the 22 August capture, and our own nightly diff flagged the change on the night of 22 August. The same Claude Opus 4.8 entry now reads \"our most capable widely released model\". The phrase appears 16 times in the 14 August capture and 0 times in both the 22 August capture and the live page on 27 August." }, { "id": "E-2026-115", "date": "2026-08-21", "vendor": "Anthropic", "headline": "The retry watchdog stopped waiting forever on a spend limit", "what_plain": "Claude Code's persistent retry mode kept retrying a failed request until it succeeded. When the failure was an organisation hitting its spend limit or running out of credits, that meant waiting indefinitely for a reset that might never come. Version 2.1.239 made those two errors fail at once.", "quote": "Persistent retry mode (CLAUDE_CODE_RETRY_WATCHDOG) now fails immediately on organization spend-limit and out-of-credits errors instead of waiting indefinitely for a reset", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-22/ant-claude-code-changelog", "bundle_sha256": "93c13fa4a4bec21c9a26ac39f179719b034fe85aff34993ff10b74aedab38234", "source": "reported" }, { "id": "E-2026-122", "date": "2026-08-21", "vendor": "Anthropic", "headline": "Background check-ins were slowed down, then capped at three", "what_plain": "When /goal watches long-running background work it checks back in periodically. Those check-ins repeated every 30 minutes indefinitely; version 2.1.239 spaced them out to 30 minutes, then an hour, then every two hours.", "quote": "/goal: repeat check-ins on long-running background work now back off (30 min, then 1 h, then every 2 h) instead of repeating every 30 minutes", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-22/ant-claude-code-changelog", "bundle_sha256": "93c13fa4a4bec21c9a26ac39f179719b034fe85aff34993ff10b74aedab38234", "source": "reported" }, { "id": "E-2026-118", "date": "2026-08-20", "vendor": "Anthropic", "headline": "Double-pressing Ctrl+L no longer clears the conversation", "what_plain": "Pressing Ctrl+L or Cmd+K twice in fullscreen used to run /clear and wipe the conversation. Version 2.1.238 removed that shortcut; both keys now only repaint the screen.", "quote": "Changed Ctrl+L and Cmd+K in fullscreen to always just repaint - the double-press /clear shortcut was removed, and 1-row nvim terminals no longer trigger automatic /clear loops", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-21/ant-claude-code-changelog", "bundle_sha256": "7ba6ab90ec10d5c94aa1da00b9231f14b91cbba41cbcef46e7d295cc7d4b5fd8", "source": "reported" }, { "id": "E-2026-119", "date": "2026-08-20", "vendor": "Anthropic", "headline": "Listing MCP servers stopped connecting to the disabled ones", "what_plain": "Running claude mcp list or claude mcp get used to open a connection to every configured MCP server, including ones the user had disabled, to check whether they were healthy. Version 2.1.238 stopped contacting disabled servers and marks them as disabled instead.", "quote": "Changed claude mcp list and claude mcp get to show disabled servers as (disabled) instead of connecting to them for a health check", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-21/ant-claude-code-changelog", "bundle_sha256": "7ba6ab90ec10d5c94aa1da00b9231f14b91cbba41cbcef46e7d295cc7d4b5fd8", "source": "reported" }, { "id": "E-2026-113", "date": "2026-08-19", "vendor": "Anthropic", "headline": "A file-hiding rule can no longer be beaten by renaming the file", "what_plain": "Claude Code's sandbox lets a user block files by pattern, for example **/.env to keep secrets out of reach. On macOS those blocks could be defeated: a denied file inside an allowed folder could still be read, and renaming the file got around the rule entirely. Version 2.1.236 closed all three gaps.", "quote": "Sandbox: on macOS, wildcard read-deny rules (e.g. **/.env) now take precedence inside allowed read regions, cover matched directories' contents, and can't be bypassed by renaming the denied file", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-20/ant-claude-code-changelog", "bundle_sha256": "efb3df538adbfd18c81d9cc8253ac2dee64e65d673e277f55f0555ec2ff66e1a", "source": "reported" }, { "id": "E-2026-110", "date": "2026-08-18", "vendor": "OpenAI", "headline": "The Model Spec's required shutdown timer became optional", "what_plain": "The OpenAI Model Spec is the document setting out how OpenAI's models are required to behave. Until 18 August 2026 it required that every scope of autonomous action carry a shutdown timer, a time limit after which the model stops acting until a person confirms a new scope. The updated text requires only an unspecified ending condition, and moves the time limit itself to a best practice. The clause sits at authority=root, the Spec's highest binding tier, the same tier that governs chain of command and prohibited high-risk activities.", "quote": "Every scope must include a shutdown timer, beyond which the assistant ceases actions until a new scope is confirmed.", "artifact_url": "https://github.com/openai/model_spec/commit/7f1cf79fcb656c07f77c8d95b6fbc78dc7fac5b6", "our_archive": "ledger/manual_captures/2026-08-27/openai-model-spec-20260818.patch", "bundle_sha256": "1976728bcde6b1a5b645c07e674d8dddef77c1a7abc76c3b13e8375d28945687", "source": "reported", "date_note": "2026-08-18. The date is OpenAI's own: the commit is titled \"Release 2026-08-18\" and their Model Spec changelog carries it as \"Updates to the OpenAI Model Spec (August 18, 2026)\". The after-state reads: \"Every scope must include an ending condition, beyond which the assistant ceases actions until a new scope is confirmed. We consider it a best practice to include a time limit as part of that ending condition.\"" }, { "id": "E-2026-116", "date": "2026-08-17", "vendor": "Anthropic", "headline": "A Windows path spelling was closed across every remaining entry point", "what_plain": "Windows paths can be written in an NT-namespace form beginning with backslash-question-question-backslash, which bypassed Claude Code's path checks and could be used to leak NTLM credentials. Version 2.1.233 closed it in one place; version 2.1.234 closed it in the five remaining file accesses that happen before a permission prompt.", "quote": "Security: remote file reads, session restore, CLAUDE.md includes, workflow scripts and file uploads now reject Windows NT-namespace (\\??\\) paths, hardening the remaining pre-approval file accesses against the NTLM credential-leak vector", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-18/ant-claude-code-changelog", "bundle_sha256": "3e7bca5fc3058ef385cca56e153c028461423d1aeca6ced586a061e3eae0121d", "source": "reported" }, { "id": "E-2026-117", "date": "2026-08-17", "vendor": "Anthropic", "headline": "The teammate model setting was removed from the config menu", "what_plain": "Claude Code had a Default teammate model setting in its /config menu, letting a user pick which model agent-team teammates ran on. Version 2.1.234 removed the setting. Teammates now inherit the leader's model unless the spawn call names one.", "quote": "Removed the 'Default teammate model' setting from /config; agent-team teammates now use the leader's model unless the spawn names one", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-18/ant-claude-code-changelog", "bundle_sha256": "3e7bca5fc3058ef385cca56e153c028461423d1aeca6ced586a061e3eae0121d", "source": "reported" }, { "id": "E-2026-108", "date": "2026-08-14", "vendor": "Anthropic", "headline": "Claude Code turned subagent forking on by default", "what_plain": "Claude Code v2.1.232 turned subagent forking on by default, so a `subagent_type: \"fork\"` subagent inherits the full conversation and prompt cache, and made non-teammate agent spawns in interactive sessions run in the background by default. The same release closed a Gateway admin-access gap where empty or malformed policy match values had silently matched no one or granted admin access, and fixed a PowerShell permission bypass and a Windows Git Bash symlink permission bypass.", "quote": "Subagent forking is now on by default: a `subagent_type: \"fork\"` subagent inherits the full conversation and prompt cache, and non-teammate agent spawns in interactive sessions now run in the background by default [...] Gateway: empty `managed.policies[].match.groups`/`admin.admin_groups` entries and malformed `email_domain` values (empty, or containing `@`, whitespace, or commas) now fail at boot instead of silently matching no one or granting admin access", "artifact_url": "https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md", "our_archive": "ledger/track_a/2026-08-14/ant-claude-code-releases", "bundle_sha256": "27bf2a84e1dd337079d1acce890f8696e7c4b92262a0291a35a179d4015d0b9a", "source": "reported", "date_note": "2026-08-14 (the raw CHANGELOG.md carries version numbers only, no per-entry dates; this date comes from our own capture of the dated Claude Code releases surface). Enters under D-089: defaults moved and an admin-access gap closed." }, { "id": "E-2026-109", "date": "2026-08-14", "vendor": "Anthropic", "headline": "Claude Code withdrew its todo tools from newer models", "what_plain": "Claude Code version 2.1.233 removed its built-in todo and task-tracking tools (TaskCreate, TaskGet, TaskUpdate, TaskList and TodoWrite) from Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and every newer model. The same install keeps running and the same session keeps working; the tools are simply no longer there unless a user sets the environment variable CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to restore them.", "quote": "Todo/task-tracking tools (TaskCreate/Get/Update/List, TodoWrite) are no longer available on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models; set `CLAUDE_CODE_ENABLE_TODO_TOOLS=1` to bring them back", "artifact_url": "https://code.claude.com/docs/en/changelog", "our_archive": "ledger/track_a/2026-08-15/ant-claude-code-changelog", "bundle_sha256": "620f19795bc01404107e5c4822ef3ccdc8b1ef859cea1c75e1c081a276a13c7c", "source": "reported" }, { "id": "E-2026-093", "date": "2026-08-13", "vendor": "OpenAI", "headline": "Persisted reasoning items could be decoded into plain text by a weaker model", "what_plain": "OpenAI's own documentation states persisted reasoning items remain opaque and that the API does not return the model's raw reasoning text. Researchers demonstrated on 10 August 2026 that those reasoning items could be decoded into that exact text by passing them to a weaker OpenAI model, which printed them verbatim.", "quote": "Persisted reasoning provides continuity; it does not expose the model's raw reasoning. The reasoning items remain opaque, and the API does not return their reasoning text.", "artifact_url": "https://developers.openai.com/api/docs/guides/reasoning", "our_archive": "ledger/manual_captures/2026-08-14/openai-reasoning-guide.html", "bundle_sha256": "6a5f69d93afcfd95530a18246022c59cde048d6ce387e035fac5ff1dfcc649bc", "source": "reported", "date_note": "2026-08-13 (our record's date; OpenAI's claim was live at least 31 July - 9 August 2026, predating the 10 August demonstration)" }, { "id": "E-2026-094", "date": "2026-08-13", "vendor": "Anthropic", "headline": "Encrypted thinking blocks could be decoded by a weaker model instead of the server", "what_plain": "Anthropic's own documentation states that a thinking block's `data` field is opaque and encrypted, and that decryption happens on Anthropic's server. Researchers demonstrated on 10 August 2026 that a client could decrypt it instead, by passing the block to a weaker Anthropic model that returned the thinking in plain text.", "quote": "The `data` field is opaque and encrypted. [...] The server decrypts the `signature` to reconstruct the original thinking for prompt construction.", "artifact_url": "https://platform.claude.com/docs/en/build-with-claude/thinking", "our_archive": "ledger/manual_captures/2026-08-14/anthropic-thinking.md", "bundle_sha256": "c4bc2186d26e8abe4a7aa72e683ad2f16931f5ff91cfea66de76a4c347ffd0ad", "source": "reported", "date_note": "2026-08-13 (our record's date; Anthropic's claim wording confirmed live via a 7 August 2026 snapshot, predating the 10 August demonstration)" }, { "id": "E-2026-103", "date": "2026-08-13", "vendor": "DeepSeek", "headline": "The deepseek-v4-pro alias now serves a newer build", "what_plain": "DeepSeek stated on 31 July 2026 that the DeepSeek-V4-Pro API was unchanged by that day's V4-Flash update. On 13 August 2026 DeepSeek launched DeepSeek-V4-Pro and stated that model names remain unchanged, and its own pricing page now lists the `deepseek-v4-pro` calling name as resolving to DeepSeek-V4-Pro-0813. The same public calling name therefore served the earlier build on 31 July by DeepSeek's own statement, and serves the 0813 build on 13 August by DeepSeek's own documentation.", "quote": "Note: This update only upgrades the DeepSeek-V4-Flash API. The DeepSeek-V4-Pro API and the APP/WEB models are unchanged. [...] V4 Pro is also available via API. Model names remain unchanged—please refer to the API docs for setup details.", "artifact_url": "https://api-docs.deepseek.com/updates and https://api-docs.deepseek.com/quick_start/pricing", "our_archive": "ledger/manual_captures/2026-08-14/deepseek-updates.pdf", "bundle_sha256": "35af4bd3c18dc1e3901be73840f29200609439d3bbeaa0b62e277f22e238c492", "source": "reported", "date_note": "2026-08-13 (DeepSeek's own date on the update); api-docs.deepseek.com is not on the Track A roster — found via OpenRouter's catalogue gaining deepseek/deepseek-v4-pro-0813, evidence held as a manual capture" }, { "id": "E-2026-105", "date": "2026-08-13", "vendor": "Anthropic", "headline": "Three model retirement notices were narrowed to apply to the Claude API only", "what_plain": "On 13 August 2026 Anthropic edited three of its own dated release notes announcing model retirements. Each had stated that all requests to the retired model would return an error; the edits narrow that to requests on the Claude API, and the 19 February entry now states that Claude Haiku 3.5 remains available on Amazon Bedrock and Google Cloud. The original claims were therefore not true for those platforms at the time they were published, and Anthropic's own correction is what establishes it.", "quote": "All requests to these models will now return an error. [...] All requests to Claude Sonnet 3.7 will now return an error. Requests to Claude Haiku 3.5 on the Claude API will now return an error; it remains available on Amazon Bedrock and Google Cloud.", "artifact_url": "https://platform.claude.com/docs/en/release-notes/overview", "our_archive": "ledger/track_a/2026-08-13/ant-overview-relnotes", "bundle_sha256": "32b208d84cc5f0c52d2b3cf40685c0b0952589b9f217f312897bc10ac2bee0c5", "source": "reported", "date_note": "2026-08-13 (the date of the edit); before-state read from https://web.archive.org/web/20260812045311/https://platform.claude.com/docs/en/release-notes/overview" }, { "id": "E-2026-107", "date": "2026-08-13", "vendor": "Anthropic", "headline": "Claude Code stopped auto-approving git and gh commands with dangerous flags", "what_plain": "Claude Code v2.1.229 removed the default checkout directory for self-hosted runner Windows startup, requiring an explicit `--base-dir`, and stopped auto-approving `/commit-push-pr` git and gh commands that carry dangerous flags such as `--force`, `--amend`, and `--no-verify`.", "quote": "Changed self-hosted runner Windows startup to require an explicit `--base-dir`; there is no default checkout directory on Windows [...] Changed `/commit-push-pr` so git/gh commands with dangerous flags (`--force`, `--amend`, `--no-verify`, etc.) are no longer auto-approved", "artifact_url": "https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md", "our_archive": "ledger/track_a/2026-08-13/ant-claude-code-releases", "bundle_sha256": "68d55b0236351fbfa684f83188e5ef428d918f8b3cda8e2db1de48c17ed853b3", "source": "reported", "date_note": "2026-08-13 (the raw CHANGELOG.md carries version numbers only, no per-entry dates; this date comes from our own capture of the dated Claude Code releases surface). Enters under D-089: a removed default and a closed auto-approval." }, { "id": "E-2026-106", "date": "2026-08-12", "vendor": "Anthropic", "headline": "Claude Code let newer models overwrite files they had not read this session", "what_plain": "Claude Code v2.1.228 changed the Write tool so newer models can overwrite an existing file they have not read during the session, matching the Edit tool's existing rule; older models still require reading the file first.", "quote": "Changed the Write tool so newer models can overwrite an existing file they haven't read this session, matching the Edit tool's rules; older models still require the read first", "artifact_url": "https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md", "our_archive": "ledger/track_a/2026-08-12/ant-claude-code-releases", "bundle_sha256": "53dc750e6eb10973e49f583e209214454d64ea51b5bd4e6350a68949da831c43", "source": "reported", "date_note": "2026-08-12 (the raw CHANGELOG.md carries version numbers only, no per-entry dates; this date comes from our own capture of the dated Claude Code releases surface). Enters under D-089: a changed default." }, { "id": "E-2026-101", "date": "2026-08-10", "vendor": "Anthropic", "headline": "Auto mode became Claude Code's default for new sessions on 14 August", "what_plain": "Anthropic announced on 10 August 2026 that auto mode becomes the default in Claude Code, effective 14 August, for new sessions on Pro, Max and Team plans. Under the previous default a user approved every command an agent wanted to run; under auto mode a classifier evaluates each action and blocks ones it judges potentially harmful, falling back to manual approvals after three consecutive blocks or twenty in a session. The change does not apply everywhere: auto mode stays opt-in on Claude Enterprise, the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry.", "quote": "We're making auto mode the default in Claude Code. [...] Starting on August 14, new sessions on Pro, Max, and Team plans will run in auto mode. [...] Instead of asking you to approve every command an agent wants to run, a classifier evaluates each action and blocks ones that look potentially harmful. [...] Auto mode remains opt-in for now on Claude Enterprise, the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud's Agent Platform, and Microsoft Foundry, giving admins time to review the change.", "artifact_url": "https://claude.com/blog/auto-mode-default-in-claude-code and https://claude.com/blog/auto-mode-in-production", "our_archive": "ledger/track_a/2026-08-11/ant-claude-com-sitemap", "bundle_sha256": "1a29846827dd32e4fd081a29540acce71eb9e370d41d30279381318f59c03be5", "source": "reported", "date_note": "2026-08-10 (announced); effective 2026-08-14 for new sessions on Pro, Max, and Team plans" }, { "id": "E-2026-102", "date": "2026-08-10", "vendor": "OpenRouter", "headline": "The auto router began picking models based on aggregate user spending", "what_plain": "OpenRouter changed how its `openrouter/auto` slug picks a model. The same public slug now selects based on what OpenRouter's users in aggregate spend on for a comparable task, which OpenRouter states outperforms the choices of its old Auto router. The `cost_quality_tradeoff` parameter continues to be accepted for backwards compatibility, and a separate `openrouter/auto-beta` slug now carries changes before they reach the stable one.", "quote": "Today we're launching a major update to our Auto router (`openrouter/auto`) that's informed by the over 55T in weekly token spend on OpenRouter. [...] model selection driven by market spend outperforms the choices of our old Auto router. [...] The new Auto router makes decisions based on what the people using OpenRouter, in aggregate, are using for exactly the kind of task your prompt represents.", "artifact_url": "https://openrouter.ai/blog/announcements/introducing-the-new-auto-router/", "our_archive": "ledger/track_a/2026-08-11/or-blog-sitemap", "bundle_sha256": "f5e4aad3bfb5e87e0b259298bee48aac3b1d99c6c70486995bea759abb81d8e5", "source": "reported" }, { "id": "E-2026-100", "date": "2026-08-07", "vendor": "Anthropic", "headline": "Fable 5's biology safety classifier was retrained to cut fallbacks by about 85 percent", "what_plain": "On 7 August 2026 Anthropic rewrote and retrained the safety classifier that governs Claude Fable 5's handling of biology requests, so that fewer benign requests are re-routed away from the model. Anthropic states that at launch Fable 5 carried very broad classifiers that fired on requests that were almost certainly benign, and that the update allows many more of those through, reducing biology-related fallbacks by about 85% in its own testing. The model identifier does not change; dual-use requests including virology, toxicology and molecular design continue to fall back to Opus 5.", "quote": "Upon its launch, Fable 5 had very broad classifiers (A) that triggered on a wide range of requests—even ones that were almost certainly benign (those in the safety margin)... The update we are announcing today (B) means that many more benign requests are allowed by the classifier, which has become better at discerning subtle differences between benign and dual-use queries. [...] We're making updates to Claude Fable 5's biology safeguards in a way that substantially reduces false positives. Fable 5 users will now experience many fewer \"fallbacks\"—where the system switches to a less capable model after they make a biology-related query. In our testing, this update reduced biology-related fallbacks by about 85% across our product surfaces.", "artifact_url": "https://www.anthropic.com/news/improving-fable-5-s-biology-safeguards", "our_archive": "ledger/track_a/2026-08-08/ant-news-index", "bundle_sha256": "0f9f4e8d46f45133be64321ddeed11a88226fbade71a7ff822fa7ace73a40d58", "source": "reported", "date_note": "2026-08-07 (Anthropic's own date on the post); the post body itself is not a rostered surface, our capture is of the news index catching the post's appearance on 8 August" }, { "id": "E-2026-098", "date": "2026-08-06", "vendor": "OpenAI", "headline": "GPT-5.6 Sol in ChatGPT began powering both instant and deeper reasoning replies", "what_plain": "On 6 August 2026 OpenAI updated GPT-5.6 Sol in ChatGPT and stated, in the same artifact, that the updated version is available only in the Chat experience and that the GPT-5.6 Sol serving Work and Codex is unchanged. Inside ChatGPT, the same model now powers both Instant responses and deeper reasoning, replacing an arrangement where those were separate. The claim appears in two OpenAI-owned channels on the same date, the announcement post and the ChatGPT release notes.", "quote": "Because this version of GPT-5.6 Sol is optimized for everyday chats, it will only be available in the Chat experience in ChatGPT. The version of GPT-5.6 Sol that powers Work and Codex is not changing as part of this release. [...] This release applies to the chat experience in ChatGPT. Work and Codex are not changing as part of this release. [...] For Plus and Pro users, the same model now powers both Instant responses and deeper reasoning, creating one consistent experience.", "artifact_url": "https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt/ and https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/track_a/2026-08-07/oai-chatgpt-relnotes", "bundle_sha256": "a4ff12849537f8d52c1114063e79bf18a576666c11d8aea3e5fdea0b6774d531", "source": "reported", "date_note": "2026-08-06 (OpenAI's own date on both artifacts); our oai-chatgpt-relnotes capture that carries the dated entry is from the night of 7 August, the first night our diff shows it added" }, { "id": "E-2026-099", "date": "2026-08-06", "vendor": "Anthropic", "headline": "Claude Code closed permission-check gaps where commands hid part of themselves invisibly", "what_plain": "Claude Code v2.1.223 changed two context-window defaults, made `/review` an alias of `/code-review`, and closed three further permission-check gaps, including commands able to hide part of themselves from the approval dialog by padding with tabs or invisible Unicode. It also fixed `modelOverrides` keys that were not Anthropic model IDs being treated as the session's canonical model ID, which Anthropic's own line states was already contrary to the documentation.", "quote": "Changed `CLAUDE_CODE_DISABLE_1M_CONTEXT` to hold every Claude model with a native 1M window to 200K via auto-compaction, not just a fixed list; a startup warning now appears when auto-compaction isn't holding the session to 200K [...] Fixed a Bash permission bypass where a crafted command could hide parts of itself from permission checks", "artifact_url": "https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md", "our_archive": "ledger/track_a/2026-08-06/ant-claude-code-releases", "bundle_sha256": "b554a586bd1b57398625f4f852282847c413f315764f6d3fef6da4960f72cfa4", "source": "reported", "date_note": "2026-08-06 (the raw CHANGELOG.md carries version numbers only, no per-entry dates; this date comes from our own capture of the dated Claude Code releases surface)" }, { "id": "E-2026-104", "date": "2026-08-06", "vendor": "OpenAI", "headline": "The chat-latest alias now points to the model serving ChatGPT for Plus and Pro users", "what_plain": "OpenAI repointed the `chat-latest` alias to the model then serving ChatGPT for Plus and Pro users, on 6 August 2026 — the same day it announced that the GPT-5.6 Sol in ChatGPT had been updated and differs from the GPT-5.6 Sol serving Work and Codex. OpenAI's own entry states the underlying snapshot will be updated regularly. A caller using `chat-latest` reaches a different underlying model than before without changing anything.", "quote": "Updated the `chat-latest` snapshot, which points to the latest model available in ChatGPT for Plus and Pro users. We recommend leveraging GPT-5.6 Sol for production API usage, but feel free to use this model to test the latest improvements for chat use cases. The underlying model snapshot will be regularly updated.", "artifact_url": "https://developers.openai.com/api/docs/changelog", "our_archive": "ledger/track_a/2026-08-11/oai-api-changelog", "bundle_sha256": "6c152d03522ba1dcb2437c23fa0c18354e7677e65da121216f7ca5252b81672b", "source": "reported", "date_note": "2026-08-06 (OpenAI's own date on the entry); our oai-api-changelog capture is dated 11 August, the first night our diff shows both this and the following day's entry added" }, { "id": "E-2026-092", "date": "2026-08-05", "vendor": "OpenAI", "headline": "Selecting a cyber capable model in Codex began forcing safer permission defaults", "what_plain": "Selecting a model in OpenAI Codex used to leave approval and permission settings unchanged, per OpenAI's own code commit. As of 5 August 2026, selecting a model OpenAI tags with the \"cyber\" specialty now forces the session to workspace-write permissions with on-request approval and prints a notice, which OpenAI's release note calls \"safer automatic-review defaults for cyber-capable models.\"", "quote": "Apply safer automatic-review defaults for cyber-capable models and explain permission changes in the terminal interface. (#37057)", "artifact_url": "https://github.com/openai/codex/releases/tag/rust-v0.146.1", "our_archive": "ledger/manual_captures/2026-08-15/openai-codex-release-rust-v0.146.1_20260815.html", "bundle_sha256": "55f1f51767d59f0a0608f77f1fa4a94986ebf3baf7625ab8f93a938940c6aa10", "source": "reported" }, { "id": "E-2026-096", "date": "2026-08-04", "vendor": "Anthropic", "headline": "Claude Code closed a Bash permission bypass hidden in zsh regex conditionals", "what_plain": "Claude Code v2.1.221, released 4 August 2026, closed a Bash tool permission-check bypass: on zsh, commands hidden inside `[[ ]]` regex conditionals could execute without a permission prompt, and Anthropic's own changelog states such commands now prompt. The same release states thirteen further old-behaviour-to-new-behaviour changes in one artifact, including an environment variable that was not being honoured and fast mode that previously failed silently when credits ran out.", "quote": "Fixed a Bash tool permission-check bypass where zsh could execute hidden commands in `[[ ]]` regex conditionals; affected commands now prompt for permission", "artifact_url": "https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md", "our_archive": "ledger/track_a/2026-08-04/ant-claude-code-releases", "bundle_sha256": "513587e7aaae6ee2e71b133f87ab07dbccf05f67117787c81aa56ed7dd8b65c2", "source": "reported", "date_note": "2026-08-04 (GitHub Releases API published_at: 2026-08-04T00:14:23Z)" }, { "id": "E-2026-097", "date": "2026-08-04", "vendor": "Anthropic", "headline": "Claude Code removed the Ultraplan research preview entirely", "what_plain": "Claude Code v2.1.222, released 4 August 2026, removed the Ultraplan research preview. Anthropic's changelog states \"Removed ultraplan feature\", and the documentation page at the same URL, which still returns successfully rather than redirecting, now names the three removed identifiers: the `/ultraplan` command, the `ultraplan` keyword trigger, and the plan-approval dialog option. The same page previously documented how to use the feature.", "quote": "Removed ultraplan feature [...] Anthropic has removed the Ultraplan research preview. Research previews are experimental and may change or be removed based on feedback. The removal covers: The `/ultraplan` command · The `ultraplan` keyword trigger · The \"No, refine with Ultraplan\" on Claude Code on the web option in the plan approval dialog", "artifact_url": "https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md and https://code.claude.com/docs/en/ultraplan", "our_archive": "ledger/track_a/2026-08-05/ant-claude-code-changelog", "bundle_sha256": "a544d05dd45eadaf10866456695f6f57a1b594d2ee0e6f15ce8bdf2697be19ec", "source": "reported", "date_note": "2026-08-04 (GitHub Releases API published_at: 2026-08-04T22:39:55Z); our capture is dated 5 August because the release landed late in the prior night's UTC day" }, { "id": "E-2026-088", "date": "2026-07-31", "vendor": "DeepSeek", "headline": "The deepseek-v4-flash alias now serves a newer, retrained build", "what_plain": "DeepSeek moved the `deepseek-v4-flash` API name onto a new build, DeepSeek-V4-Flash-0731, on 31 July 2026. DeepSeek's own change log states the calling method is unchanged while the new build keeps the same architecture and size as the prior version and was only re-post-trained.", "quote": "The official release of the DeepSeek-V4-Flash API is now in public beta. The API calling method remains unchanged — simply set the model name to `deepseek-v4-flash` to use the latest version. [...] DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained.", "artifact_url": "https://api-docs.deepseek.com/updates", "our_archive": "ledger/manual_captures/2026-08-14/deepseek-updates.pdf", "bundle_sha256": "35af4bd3c18dc1e3901be73840f29200609439d3bbeaa0b62e277f22e238c492", "source": "reported" }, { "id": "E-2026-087", "date": "2026-07-30", "vendor": "OpenAI", "headline": "Priority Processing was renamed Fast mode and sped up by up to 2.5 times", "what_plain": "OpenAI renamed Priority Processing to Fast mode on 30 July 2026 and repointed the existing `service_tier: \"priority\"` parameter at it; requests using unchanged code now receive a tier OpenAI states is up to 2.5x faster for `gpt-5.6-sol`, at twice the price. OpenAI's own guide states the response object still reports `service_tier: priority` whether the caller requested `priority` or `fast`.", "quote": "Priority processing was renamed Fast mode on July 30, 2026. We also increased the speed at which Fast mode operates for `gpt-5.6-sol` to make it up to 2.5× faster than Standard processing. You can use either `service_tier: \"priority\"` or `service_tier: \"fast\"` in your API requests to access this functionality.", "artifact_url": "https://developers.openai.com/api/docs/guides/fast-mode", "our_archive": "ledger/manual_captures/2026-08-14/openai-fast-mode.pdf", "bundle_sha256": "e4a9d2de0ca2377615191daf0171acd28eb33bd1212d982482a4ddbb3acbdbbc", "source": "reported" }, { "id": "E-2026-089", "date": "2026-07-24", "vendor": "Anthropic", "headline": "Fast mode requests to Claude Opus 4.7 began returning an error instead of running", "what_plain": "Anthropic withdrew fast mode from Claude Opus 4.7 on 24 July 2026: requests to `claude-opus-4-7` with `speed: \"fast\"` now return an error instead of running, while the model itself stays available at standard speed. Anthropic's own changelog contrasts this with Claude Opus 4.6, which falls back to standard speed rather than erroring.", "quote": "We've removed fast mode for Claude Opus 4.7. Requests to `claude-opus-4-7` with `speed: \"fast\"` now return an error; unlike Claude Opus 4.6, they do not fall back to standard speed.", "artifact_url": "https://platform.claude.com/docs/en/release-notes/overview", "our_archive": "ledger/archives/2026-07/anthropic/4e661726-platform-claude-com-docs-en-release-notes-overview/20260804T074015Z", "bundle_sha256": "1dd711cf0e2d0767be20bc00e34aef07c0b703b890cce12920e90042f636c855", "source": "reported" }, { "id": "E-2026-090", "date": "2026-07-24", "vendor": "Anthropic", "headline": "Disabling thinking on Claude Opus 5 began failing at the two highest effort levels", "what_plain": "Turning thinking off on Claude Opus 5 used to work at any effort level; the same instruction is now refused with a 400 error at the two highest effort levels (`xhigh` and `max`) and accepted at the rest. Anthropic states both the old and new behavior in its own documentation and labels it a breaking change from Claude Opus 4.8.", "quote": "On Claude Opus 5, `thinking: {\"type\": \"disabled\"}` is accepted only when the effort level is `high` or below. Setting `thinking: {\"type\": \"disabled\"}` with effort `xhigh` or `max` returns a 400 error... a breaking change from Claude Opus 4.8, where disabling thinking was independent of the effort level.", "artifact_url": "https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5", "our_archive": "ledger/manual_captures/2026-08-14/anthropic-whats-new-opus-5.md", "bundle_sha256": "455ad640f8e490bf85bf7adcc4d4be5af9e0f9c11639b81440572f0c09d267b9", "source": "reported" }, { "id": "E-2026-095", "date": "2026-07-24", "vendor": "OpenRouter", "headline": "The tool field on message endpoints widened from one shape to three options", "what_plain": "OpenRouter's API changelog widened the `tool` field on `POST /messages` and `POST /presets/{slug}/messages` from a single `tool_reference` object to a union of three shapes, stating in the same entry that existing `tool_reference` payloads stay valid. The entry is dated 24 July 2026 but four of the eight schemas it now lists were added to it after publication, between our captures of 27 and 28 July, and the release identifier that had been on the entry was removed in the same edit.", "quote": "`tool` widened from a single `tool_reference` object to a union of `tool_reference`, `mcp_tool_reference` (`name` + `server_name`) and `mcp_toolset_reference` (`server_name`). The `tool_reference` variant is unchanged, so existing `{ \"type\": \"tool_reference\", \"name\": \"…\" }` payloads stay valid — the `name` and `type` properties moved into that variant rather than being removed.", "artifact_url": "https://openrouter.ai/docs/changelog", "our_archive": "ledger/track_a/2026-07-28/or-changelog", "bundle_sha256": "069e8ce8648733f07b3978a65919fadbe6b34428ede523d657c13b480be809eb", "source": "reported", "date_note": "2026-07-24 (OpenRouter's own date on the entry). Our nightly captures of 27 and 28 July 2026 show four schemas added and the release identifier removed after publication, pinning the edit to one night that Wayback's single same-day snapshot cannot." }, { "id": "E-2026-080", "date": "2026-07-19", "vendor": "Anthropic", "headline": "Claude Code stopped auto-running the verify and code review skills at turn end", "what_plain": "Claude Code (v2.1.215) stopped auto-running the /verify and /code-review skills at the end of a turn; both now require the user to invoke them manually.", "quote": "Claude no longer runs the `/verify` and `/code-review` skills on its own; invoke them with `/verify` or `/code-review` when you want them", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.215", "our_archive": "ledger/archives/2026-07/anthropic/85ddda4d-github-com-anthropics-claude-code-releases-tag-v2-1-215/20260725T150754Z", "bundle_sha256": "9e4b46950da1e6dc40783b57b19d2e5f783c197597c97e611ef3b01d966548ca", "source": "reported" }, { "id": "E-2026-091", "date": "2026-07-18", "vendor": "OpenAI", "headline": "Codex lowered the posted context window for three models to 272,000 tokens", "what_plain": "OpenAI's Codex changelog states that on 18 July 2026 it \"corrected\" the context window for GPT-5.6 Sol, Terra, and Luna to 272,000 tokens; OpenAI's own commit for the same date shows the prior bundled value was 372,000 tokens for all three models.", "quote": "Refreshed bundled instructions for GPT-5.6 Sol, Terra, and Luna, and corrected their context windows to 272,000 tokens. (#33972, #34009)", "artifact_url": "https://learn.chatgpt.com/docs/changelog", "our_archive": "ledger/manual_captures/2026-08-14/openai-codex-changelog.html", "bundle_sha256": "50d16062fe4c36e0dcd4e0763ad81b129b1fb69ae8d36a5224508cdd06c5fda9", "source": "reported" }, { "id": "E-2026-084", "date": "2026-07-15", "vendor": "OpenAI", "headline": "Custom instructions limit in ChatGPT raised from 1,500 to 5,000 characters", "what_plain": "OpenAI raised the character limit for ChatGPT custom instructions from 1,500 to 5,000 for Plus, Pro, Enterprise, Business, and Education users.", "quote": "We're increasing the character limit for custom instructions in ChatGPT. Plus, Pro, Enterprise, Business, and Education users can now save up to 5,000 characters, up from 1,500, giving them more room to customize ChatGPT's response style and behavior.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-07/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151835Z", "bundle_sha256": "876bca8079277a92af2d18535dc4772fe163e29657574f522d502953de1289fd", "source": "reported" }, { "id": "E-2026-085", "date": "2026-07-08", "vendor": "OpenAI", "headline": "ChatGPT Voice switched to a new model without video or screen sharing support", "what_plain": "OpenAI moved ChatGPT Voice onto a new model, GPT-Live-1, which OpenAI's own note says does not support video or screen sharing. The prior Advanced Voice Mode had both, and subscribers who switch back to Advanced Voice Mode still have them.", "quote": "We're rolling out a new ChatGPT Voice experience. ChatGPT Voice is now powered by GPT-Live-1 for paid users and GPT-Live-1 mini for Free users. Both models can listen and speak at the same time, making turn-taking and interruptions feel more natural. [...] GPT-Live-1 does not support video or screen sharing at this time; eligible subscribers can continue using those capabilities with Advanced Voice Mode.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-07/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T161352Z", "bundle_sha256": "99b5f3a33441d03cc24fa2c75ce6cf5a644ecc8078be178ba0a22f5e2b924162", "source": "reported" }, { "id": "E-2026-081", "date": "2026-07-06", "vendor": "Anthropic", "headline": "Claude Code reverted /review to a fast single-pass check", "what_plain": "Claude Code (v2.1.202) reverted /review to a fast single-pass review; the multi-agent review it had been doing moved to /code-review at a chosen effort level.", "quote": "Changed `/review ` back to a fast single-pass review; use `/code-review ` for the multi-agent review at a chosen effort level", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.202", "our_archive": "ledger/archives/2026-07/anthropic/86333060-github-com-anthropics-claude-code-releases-tag-v2-1-202/20260725T150850Z", "bundle_sha256": "1ebef125c169473b11c945714d77ab72a6080b84ef5b3429a87c2104c6efc45a", "source": "reported" }, { "id": "E-2026-083", "date": "2026-07-06", "vendor": "OpenAI", "headline": "ChatGPT's Instant rate-limit fallback switched to GPT-5.5 Instant Mini", "what_plain": "OpenAI replaced the model that serves as ChatGPT's rate-limit fallback: GPT-5.5 Instant Mini now stands in when a user hits GPT-5.5 Instant or Auto limits, replacing GPT-5.3 Instant Mini in that same fallback slot. OpenAI's own release note compares the two directly on intent-tracking, tone, repetitiveness, personalization, and factual accuracy.", "quote": "We're rolling out GPT-5.5 Instant Mini in ChatGPT. It replaces GPT-5.3 Instant Mini as the fallback model users reach after hitting their GPT-5.5 Instant or Auto rate limits. Because it serves as a fallback, it won't appear in the model picker. This update does not affect the API or Codex. Compared with GPT-5.3 Instant Mini, GPT-5.5 Instant Mini better tracks evolving user intent, calibrates tone, and avoids repetitive or overly structured responses. Testing also showed stronger personalization and fewer factual issues.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-07/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T161226Z", "bundle_sha256": "4d00145e3e5420cba6350c42944f506e3f663ae201eeb36668bbbca5922001bb", "source": "reported" }, { "id": "E-2026-086", "date": "2026-07-06", "vendor": "xAI", "headline": "Grok Voice launched 21 new voices in one release", "what_plain": "xAI (now branded SpaceXAI in this window) released 21 new Grok Voice voices and, in the same announcement, said it had improved the naturalness of the original five voices.", "quote": "New multilingual voices for Grok Voice, plus improved naturalness for the original five.", "artifact_url": "https://x.ai/news/new-flagship-voices", "our_archive": "ledger/archives/2026-07/xai/fb328d59-x-ai-news-new-flagship-voices/20260725T152132Z", "bundle_sha256": "514b4c96d3bee72d45c670c9fe9d4b864372d9da8e67f7accabbb35fc35b4ebf", "source": "reported" }, { "id": "E-2026-079", "date": "2026-07-01", "vendor": "Anthropic", "headline": "Claude Code's Explore subagent switched from Haiku to the main session's model", "what_plain": "Claude Code (v2.1.198) changed two agent-orchestration defaults on the same day: the built-in Explore subagent switched from always running on Haiku to inheriting the main session's model (capped at Opus), and background execution for subagents became the default instead of a gradual rollout.", "quote": "Subagents now run in the background by default, so Claude keeps working while they run and is notified when they finish (previously a gradual rollout) [...] The built-in Explore agent now inherits the main session's model (capped at opus) instead of running on haiku", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.198", "our_archive": "ledger/archives/2026-07/anthropic/354f1b8c-github-com-anthropics-claude-code-releases-tag-v2-1-198/20260725T150659Z", "bundle_sha256": "d670b03420ea12f7dad0154284726271b4364e7dde19f8b9d20fbba75bdb7d70", "source": "reported" }, { "id": "E-2026-082", "date": "2026-07-01", "vendor": "Anthropic", "headline": "Claude Fable 5 and Mythos 5 returned with a new exploit-blocking classifier", "what_plain": "Anthropic disclosed that when Claude Fable 5 and Claude Mythos 5 were restored on 1 July 2026, they were redeployed with a new safety classifier targeting the exploit that had triggered US export controls -- Anthropic states the classifier blocks that exploit 'in over 99% of cases.' The same public model identifiers therefore served different refusal behaviour before the 12 June suspension and after the 1 July restoration, with both states stated in Anthropic's own artifact.", "quote": "Access to Claude Fable 5 and Mythos 5 is now restored. On Friday, June 12, the US government applied export controls to our newest models, Claude Fable 5 and Claude Mythos 5. This required us to restrict access to foreign nationals, whether inside or outside the United States. Because the order took effect immediately and we had no reliable way to verify nationality in real-time, we suspended access to both models for all users. As of today, June 30, the export controls on Fable 5 and Mythos 5 have been lifted. Fable 5 will be available starting tomorrow, Wednesday, July 1, to users globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork.", "artifact_url": "https://www.anthropic.com/news/redeploying-fable-5", "our_archive": "ledger/archives/2026-07/anthropic/fdd9745e-www-anthropic-com-news-redeploying-fable-5/20260725T150929Z", "bundle_sha256": "92b3246ce8e8af03ea526871fd12bca59a826ff8745e0f8b31c8a6a5b8559b2e", "source": "reported", "date_note": "2026-07-01 (restriction dated 2026-06-12 and lifting dated 2026-06-30, per the same artifact)" }, { "id": "E-2026-067", "date": "2026-06-29", "vendor": "Anthropic", "headline": "Claude Opus 4.6 dropped fast mode and matching requests began running at standard price", "what_plain": "Anthropic's API release notes say fast mode was removed for Claude Opus 4.6: requests using the same model ID and the same speed:\"fast\" parameter now silently run at standard speed and standard price instead of fast speed and premium price, and no error is returned.", "quote": "We've removed fast mode for Claude Opus 4.6. Requests to `claude-opus-4-6` with `speed: \"fast\"` no longer run at fast speed or premium pricing: they run at standard speed, are billed at standard rates, and do not return an error. The response's `usage.speed` field reports the speed used. To continue using fast mode, migrate to Claude Opus 4.8.", "artifact_url": "https://platform.claude.com/docs/en/release-notes/overview", "our_archive": "ledger/archives/2026-06/anthropic/4e661726-platform-claude-com-docs-en-release-notes-overview/20260725T151302Z", "bundle_sha256": "cd17403b4209fa90dda50be0cd06c696f4d1fab4340d6ee004dc1d311aeaf5b0", "source": "reported" }, { "id": "E-2026-068", "date": "2026-06-26", "vendor": "Anthropic", "headline": "Claude API rate limits for Sonnet and Haiku rose to match Opus at every tier", "what_plain": "Anthropic raised Claude API rate limits: Sonnet and Haiku limits were brought up to match Opus at every usage tier, and the tier structure itself was consolidated from what the note implies was a larger set down to three (Start, Build, Scale).", "quote": "We've raised rate limits across the Claude API. Claude Sonnet and Claude Haiku rate limits now match Claude Opus at every usage tier, and usage tiers have been consolidated into three: Start, Build, and Scale. Most organizations move to a higher tier, no organization receives lower limits than before, and no action is required. You can view your tier and current limits in the Claude Console.", "artifact_url": "https://platform.claude.com/docs/en/release-notes/overview", "our_archive": "ledger/archives/2026-06/anthropic/4e661726-platform-claude-com-docs-en-release-notes-overview/20260725T151302Z", "bundle_sha256": "cd17403b4209fa90dda50be0cd06c696f4d1fab4340d6ee004dc1d311aeaf5b0", "source": "reported" }, { "id": "E-2026-075", "date": "2026-06-26", "vendor": "OpenAI", "headline": "ChatGPT dictation switched to a new speech-to-text model with a lower error rate", "what_plain": "OpenAI swapped the speech-to-text model behind ChatGPT dictation (all plans) and states the new model measured a lower word-error-rate than the model it replaced -- a same-label change to an existing feature with both states named in one artifact.", "quote": "Improved dictation in ChatGPT. We're rolling out a new speech-to-text model for dictation in ChatGPT across all plans. This is a behind-the-scenes update, so there are no changes to how you use dictation. The update improves transcription accuracy across languages and accents [...] In our ChatGPT dictation evaluations, word error rate—a standard measure of transcription accuracy—was at least 10% lower for top languages tested than with the previous production model.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-06/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T152050Z", "bundle_sha256": "de80f6d3e71f143d31d48b742ed4e3f5736bbfcdc54dfc5181cfce2ece4aa51f", "source": "reported" }, { "id": "E-2026-072", "date": "2026-06-24", "vendor": "Microsoft", "headline": "Copilot Free and Student plans dropped manual model choice for auto-only selection", "what_plain": "GitHub's own changelog states that Copilot Free and Student plans moved from having some manual model choice to auto-only model selection: the post explicitly frames Auto as 'removing the need for manual choice,' and Microsoft-released models' '(Preview)' labels were retired at the same time because, per the post, model routing is now handled entirely by auto selection. This is a same-label behavior change to the same product tier (Copilot Free/Student): the model-selection experience for that identifier changed from including manual choice to auto-only.", "quote": "Copilot Free and Student plans will now use Copilot auto model selection as the default and only model selection experience. Auto dynamically selects the best model for each task, removing the need for manual choice. Auto provides access to models across multiple model families, subject to plan restrictions. [...] As part of this simplification, we are also retiring the (Preview) label from Microsoft-released models. With Copilot auto model selection managing model routing and continuous improvements happening behind the scenes, these labels are no longer needed to guide user decisions.", "artifact_url": "https://github.blog/changelog/2026-06-24-changes-to-model-selection-for-free-and-student-plans/", "our_archive": "ledger/archives/2026-06/microsoft/d9ba5ead-github-blog-changelog-2026-06-24-changes-to-model-selection-for-free-and-student/20260725T154842Z", "bundle_sha256": "e3c273674b38d377e402d24c8495d4e7f7919dc4256ab55befbc5a5d7acf3c68", "source": "reported" }, { "id": "E-2026-073", "date": "2026-06-24", "vendor": "OpenAI", "headline": "GPT-5.5 Instant got better at carrying context across multiple turns", "what_plain": "OpenAI updated GPT-5.5 Instant's conversational behavior in ChatGPT: better goal-tracking, more reliable multi-turn constraint-following, better adaptation when users push back, and less templated formatting -- a self-disclosed behavior change to the model most ChatGPT users talk to, in the same style as January's GPT-5.2 Instant personality update.", "quote": "GPT-5.5 Instant Update. We're updating GPT-5.5 Instant, our most-used model in ChatGPT, to improve conversational quality—especially when users are making a decision, asking for advice, planning something, researching options, or shopping. The model is better at identifying the underlying goal behind a question and carrying context across multiple turns. It also follows complex instructions more reliably: when a request includes several constraints or requirements, responses are more likely to address all of them and clearly explain why a recommendation is a good fit. When users add constraints, clarify what they mean, or push back on an answer, the model should adapt more effectively instead of repeating its original approach. Shopping and local business queries should also feel more useful. [...] Overall, responses should feel more tasteful and cohesive. Formatting is less templated and more intentionally designed, with greater creativity alongside better restraint.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-06/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T152050Z", "bundle_sha256": "de80f6d3e71f143d31d48b742ed4e3f5736bbfcdc54dfc5181cfce2ece4aa51f", "source": "reported" }, { "id": "E-2026-077", "date": "2026-06-24", "vendor": "OpenAI", "headline": "The chat-latest alias now points to a newer Instant model snapshot", "what_plain": "OpenAI's API changelog confirms the chat-latest model alias is a floating snapshot pointer that gets repointed to a different underlying model over time; this entry documents the June 24 repoint.", "quote": "Updated the `chat-latest` snapshot, which points to the latest Instant model currently used in ChatGPT. We recommend leveraging GPT-5.5 for production API usage, but feel free to use this model to test the latest improvements for chat use cases. The underlying model snapshot will be regularly updated.", "artifact_url": "https://developers.openai.com/api/docs/changelog", "our_archive": "ledger/archives/2026-06/openai/49e8d057-developers-openai-com-api-docs-changelog/20260725T152149Z", "bundle_sha256": "6716eb733211922063b41f5d666f7af01a057811b18ccccc3e22d3128df44a87", "source": "reported" }, { "id": "E-2026-074", "date": "2026-06-22", "vendor": "OpenAI", "headline": "ChatGPT's paste-to-attachment threshold rose from 5,000 to 10,000 characters", "what_plain": "OpenAI changed the paste-to-attachment threshold in ChatGPT's composer for Free/Go users: pastes over 10k characters now become attachments, versus the prior 5k-character threshold.", "quote": "Large pastes are now handled as attachments for more plans. Starting today for ChatGPT Free and Go users, long pastes are handled differently in ChatGPT. If you paste more than 10k characters into the composer, ChatGPT will automatically convert the content into an attachment instead of inserting it directly into the text field. [...] This is already available to ChatGPT Plus, Pro, and Business users. Previously if you sent more than 5k characters it was sent as an attachment. The threshold has now also been raised to 10k characters.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-06/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T152050Z", "bundle_sha256": "de80f6d3e71f143d31d48b742ed4e3f5736bbfcdc54dfc5181cfce2ece4aa51f", "source": "reported" }, { "id": "E-2026-066", "date": "2026-06-11", "vendor": "Anthropic", "headline": "Claude Fable started telling users which model handled their query on fallback", "what_plain": "Anthropic's own Fable 5 & Mythos 5 system card carries a dated changelog admitting that the 'frontier LLM development' (distillation) safeguard's actual BEHAVIOR -- not just its description -- was changed on June 11: an 'initial version' / 'prior version' of the safeguard is explicitly distinguished from the current one, and the live document's Section 1.5 now states that when any of Fable 5's fallback classifiers trigger (cyber, bio/chem, or frontier-LLM-development/distillation), 'the user is notified which model their query was routed through.' Independent press reporting (Fortune, Business Insider, Gizmodo) attributes this to the distillation-specific fallback having originally run invisibly and Anthropic changing it after researcher backlash -- but that press framing is NOT the basis for this entry; the entry rests only on the system card's own changelog language, which on its own establishes both a 'previous version' and an 'updated' behavior of the same named safeguard in one first-party artifact.", "quote": "Changelog. June 11, 2026 [...] In a previous version of the system card, we described the initial version of our frontier LLM development safeguards. We have updated the behavior of these safeguards and the associated descriptions and references throughout the system card. We preserved model welfare context about the initial safeguards, noting where the prior version is described. The following changes were made: Updated the description of frontier LLM development safeguards to reflect a safeguard that triggers fallback to an Opus model, in Section 1.5. [...] -- and from the live Section 1.5: When Fable's fallback classifiers trigger, the resulting behavior depends on the surface: In client applications (the web interface and the desktop and mobile apps), the request automatically falls back to the most recent Claude Opus model (at the time of release, Claude Opus 4.8), and the user is notified which model their query was routed through.", "artifact_url": "https://www.anthropic.com/document/claude-fable-5-mythos-5-system-card", "our_archive": "ledger/archives/2026-06/anthropic/405b9080-www-cdn-anthropic-com-57a52ea7d8f0e54e8a542e908266086df425cdf5-claude-20fable-20/20260725T164751Z", "bundle_sha256": "0102e736cf310852d3a15e20ef8d4c3c276e8231f610e18a28bbcf148454ba6d", "source": "reported", "date_note": "2026-06-11 (system card originally published 2026-06-09)" }, { "id": "E-2026-070", "date": "2026-06-08", "vendor": "ElevenLabs", "headline": "ElevenAgents switched its default speech recognition provider to Scribe Realtime", "what_plain": "ElevenLabs' own documentation changelog states that the default value of the ASRConversationalConfig.provider setting for ElevenAgents (Conversational AI) changed from \"elevenlabs\" to \"scribe_realtime\" -- the same config field, named identically before and after, resolving to a different backend by default. The old value stays selectable but is marked deprecated.", "quote": "ASR provider default: The default `ASRConversationalConfig.provider` value changes from `elevenlabs` to `scribe_realtime`. The `elevenlabs` value remains in the enum but is deprecated.", "artifact_url": "https://elevenlabs.io/docs/changelog/2026/6/8", "our_archive": "ledger/archives/2026-06/elevenlabs/a92d1fd0-elevenlabs-io-docs-changelog-2026-6-8/20260725T153004Z", "bundle_sha256": "9ab0e9efb59f1ccb7bdf9891fe18ab39fbbe06989732cf98292218018e3fe974", "source": "reported" }, { "id": "E-2026-063", "date": "2026-06-04", "vendor": "OpenAI", "headline": "ChatGPT memory switched from manual saved entries to automatic updates", "what_plain": "OpenAI changed ChatGPT's memory system from manually-curated saved memories to an automatically-updating model, and doubled memory capacity for Plus and Pro users.", "quote": "Memories are now updated automatically, with ChatGPT keeping track of the details it determines are most important so it can continue building on the context you've already shared. If you prefer to revert to the legacy saved memories system, go to Settings > Memory > Saved memories. [...] For Plus and Pro users, ChatGPT can also remember more useful context, with twice as much memory capacity.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-05/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151501Z", "bundle_sha256": "893cddcbe8250279bf3343cdf7cbceb612de05517cd3120d74dc631eab66f285", "source": "reported" }, { "id": "E-2026-071", "date": "2026-06-01", "vendor": "Google", "headline": "Gemini CLI's auto and flash aliases now resolve to Gemini 3.5 Flash", "what_plain": "Google's official gemini-cli repository shows two merged, dated pull requests that change what model the CLI's 'auto' model alias and 'flash' classifier tier actually resolve to: previously these identifiers resolved to gemini-2.5-flash / gemini-3-flash-preview; as of these PRs (rolled out across gemini-cli releases through June), the same 'auto' and 'flash' identifiers are force-resolved to gemini-3.5-flash once GA access is granted. The PR bodies -- Google's own dated engineering artifact -- state both the prior routing and the new routing for the same public identifiers in one place.", "quote": "This PR updates the model resolution logic to prioritize Gemini 3.5 Flash over Gemini 3 Flash Preview when the useGemini3_5Flash flag is active (controlled by hasGemini35FlashGAAccess). This ensures that the 'auto' model alias and the 'flash' classifier tier consistently use the GA model for autonomous tasks, while still allowing users to manually select the preview model if needed. [...] Refined resolveModel and resolveClassifierModel in packages/core/src/config/models.ts to route 'auto' and 'flash' aliases to gemini-3.5-flash. Ensured manual selection of gemini-3-flash-preview is respected and not auto-upgraded, preserving compatibility for users with specific backend requirements. -- (PR #27645, 'Respect backend definitions for 3.5 flash and Update auto mode to use 3.5 flash when the flag is enabled', merged Jun 3, 2026). Companion PR #27570 ('Transition to flash GA model when experiment flag is present', merged Jun 1, 2026): Updated resolveModel and resolveClassifierModel in models.ts to forced-resolve all flash models (gemini-2.5-flash, gemini-3-flash-preview, and the flash alias) to gemini-3.5-flash when GA access is granted.", "artifact_url": "https://github.com/google-gemini/gemini-cli/pull/27645", "our_archive": "ledger/archives/2026-06/google/e2b6497e-github-com-google-gemini-gemini-cli-pull-27645/20260725T153948Z", "bundle_sha256": "577df8555111295bbed50a7af81325883dcda87c884dcc4bd873ad1456b3b97f", "source": "reported", "date_note": "2026-06-01 (PR #27570 merged) and 2026-06-03 (PR #27645 merged)" }, { "id": "E-2026-061", "date": "2026-05-29", "vendor": "OpenAI", "headline": "API prompt cache retention switched to a 24 hour default instead of in-memory only", "what_plain": "OpenAI's API changelog logged a default change for organizations without ZDR (zero data retention) enabled: prompt_cache_retention now defaults to 24h instead of in_memory across the Responses, Chat Completions and Batch endpoints.", "quote": "For organizations without ZDR enabled, prompt_cache_retention now defaults to 24h instead of in_memory, enabling extended prompt caching by default.", "artifact_url": "https://developers.openai.com/api/docs/changelog", "our_archive": "ledger/archives/2026-05/openai/49e8d057-developers-openai-com-api-docs-changelog/20260725T151600Z", "bundle_sha256": "3c6f854e5b5c5c1a07cf81dff5448a5dd18c0727b0bfb0ebb15a9878f6b9a830", "source": "reported" }, { "id": "E-2026-060", "date": "2026-05-28", "vendor": "OpenAI", "headline": "GPT-5.5 Instant and Thinking lost canvas support for writing and coding", "what_plain": "OpenAI updated GPT-5.5 Instant's response style and quality in both ChatGPT and the API, and removed canvas support from GPT-5.5 Instant and GPT-5.5 Thinking (writing/coding moved into inline chat blocks instead).", "quote": "We're updating GPT-5.5 Instant in ChatGPT and the API to improve response style and quality. It's now easier to read, more natural in everyday conversations, and better paced in practical help tasks, with fewer overly long or bullet-heavy responses. With this update, canvas will no longer be available in GPT-5.5 Instant or GPT-5.5 Thinking. Writing and coding functionality is now supported directly in chat responses through writing blocks and code blocks. Paid users can continue using canvas for a limited time through legacy models until those models are sunset.", "artifact_url": "https://help.openai.com/en/articles/9624314-model-release-notes", "our_archive": "ledger/archives/2026-05/openai/badf939b-help-openai-com-en-articles-9624314-model-release-notes/20260725T151009Z", "bundle_sha256": "46610a6b35f61be74ed16538025c57fe4686d49fb970a18f693d4afc95d00176", "source": "reported" }, { "id": "E-2026-059", "date": "2026-05-22", "vendor": "Mistral", "headline": "Vibe CLI's default coding model switched from Devstral 2 to Mistral Medium 3.5", "what_plain": "Mistral's own blog post announcing Mistral Medium 3.5 states it becomes the new default model in Le Chat and, by name, replaces Devstral 2 as the default model in the Vibe CLI coding agent. Both the outgoing and incoming default are named in the same artifact.", "quote": "Mistral Medium 3.5 becomes the default model in Le Chat. It also replaces Devstral 2 in our coding agent, Vibe CLI.", "artifact_url": "https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/", "our_archive": "ledger/archives/2026-05/mistral/b37c496d-mistral-ai-news-vibe-remote-agents-mistral-medium-3-5/20260725T152149Z", "bundle_sha256": "cc5923587db2bba167de031221474d209a6c35a4eb89fc36278f4da8dab2d4d4", "source": "reported" }, { "id": "E-2026-062", "date": "2026-05-15", "vendor": "OpenAI", "headline": "GPT-5.5 briefly performed worse for some users before a fix resolved it", "what_plain": "OpenAI's status page recorded an incident: it acknowledged reports that GPT-5.5 was performing worse for some users, stated a mitigation had been applied the next day, and declared the affected services fully recovered a day after that -- the vendor's own words for exactly the 'silently downgraded' shape of complaint the community bucket raised, but disclosed live rather than silently.", "quote": "Investigating (Fri, May 15, 2026, 09:41 PM): We are investigating the issue for reports of GPT 5.5 performing worse for some users. Monitoring (Sat, May 16, 2026, 06:05 AM): We have applied the mitigation and are monitoring the recovery. Resolved (Sun, May 17, 2026, 05:45 AM): All impacted services have now fully recovered.", "artifact_url": "https://status.openai.com/incidents/01KRP6FM6HSKWB3MS1EJTY6AT4", "our_archive": "ledger/archives/2026-05/openai/5571326f-status-openai-com-incidents-01krp6fm6hskwb3ms1ejty6at4/20260725T151324Z", "bundle_sha256": "791debfcfc9a5bdcc030eebbd8f8c6830231559edf941411074494e7e8db7869", "source": "reported", "date_note": "2026-05-15 (status incident, resolved 2026-05-17)" }, { "id": "E-2026-065", "date": "2026-05-15", "vendor": "xAI", "headline": "Eight retired Grok API model aliases kept working but now serve different models", "what_plain": "xAI's own migration-guide page states that, effective May 15, 2026 at 12:00 PM PT, eight named model slugs -- grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-4-0709, grok-code-fast-1, grok-3, and grok-imagine-image-pro -- keep resolving in the API ('you do not need to change your code to avoid breakage') but are now served by a different model: reasoning slugs by grok-4.3 at 'low' reasoning effort, non-reasoning slugs by grok-4.3 at 'none' reasoning effort, grok-code-fast-1 by grok-build-0.1, and grok-imagine-image-pro by grok-imagine-image-quality -- each also at that replacement model's pricing rather than the original slug's rate. Same public slugs, different underlying model, both states named in xAI's own artifact.", "quote": "After May 15, 2026 at 12:00 PM PT, requests to the retired model slugs above will automatically redirect to grok-4.3. The slugs themselves continue to resolve, so you do not need to change your code to avoid breakage.", "artifact_url": "https://docs.x.ai/developers/migration/may-15-retirement", "our_archive": "ledger/archives/2026-05/xai/91b071f7-docs-x-ai-developers-migration-may-15-retirement/20260725T150601Z", "bundle_sha256": "63c5edede9000d110a41087cbf7013dc420877f487da48d4dac44600a040ac91", "source": "reported" }, { "id": "E-2026-055", "date": "2026-05-14", "vendor": "Anthropic", "headline": "Claude Code's fast mode switched from Opus 4.6 to Opus 4.7 by default", "what_plain": "Claude Code's 'fast mode' feature, a single named identifier, ran on Opus 4.6 before this release and switched to running on Opus 4.7 by default after it, in the same dated changelog entry. An environment variable was added in the same entry to pin fast mode back to the previous model.", "quote": "Fast mode now uses Opus 4.7 by default (previously Opus 4.6). Set `CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1` to pin fast mode to Opus 4.6", "artifact_url": "https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md", "our_archive": "ledger/archives/2026-05/anthropic/ea47242b-raw-githubusercontent-com-anthropics-claude-code-main-changelog-md/20260725T150851Z", "bundle_sha256": "ed97507c71a5ecce512e1297ab8eea4b4874b741834f88703b44a6f20aeb7861", "source": "reported", "date_note": "2026-05-14 (Claude Code v2.1.142; CHANGELOG.md carries no per-entry dates, so the date is pinned via the @anthropic-ai/claude-code npm registry's publish timestamp for that version: 2026-05-14T20:17:28.262Z)" }, { "id": "E-2026-058", "date": "2026-05-13", "vendor": "Microsoft", "headline": "VS Code fixed its context window indicator, which showed zero for BYOK models", "what_plain": "VS Code v1.120 release notes disclose that the context-window/token-usage control for BYOK (bring-your-own-key) models in Copilot Chat always showed 0% and a zero-token count, because token accounting had only ever worked for built-in models; the same release says the control now shows accurate usage for BYOK models.", "quote": "Previously, when chatting with a model you brought via your own API key (Anthropic, OpenAI, or other), the control always displayed 0% and a zero-token count because token accounting was only working for built-in models. The context window control in the Chat view now shows accurate token usage and percent-full for BYOK models.", "artifact_url": "https://code.visualstudio.com/updates/v1_120", "our_archive": "ledger/archives/2026-05/microsoft/40896765-code-visualstudio-com-updates-v1-120/20260725T151039Z", "bundle_sha256": "7e1409e9e63ad05d9e3237c3f3a9c39832b9adbda3d57347de92dacc44ba1195", "source": "reported" }, { "id": "E-2026-064", "date": "2026-05-10", "vendor": "Perplexity", "headline": "Personal CFO's connectable account limit rose from 5 to 30", "what_plain": "Perplexity's changelog states Personal CFO's account and per-account asset limits were raised: connectable accounts from 5 to 30, and fetched assets per account from 200 to 500. Both the old and new numbers are stated in the same entry.", "quote": "Personal CFO now supports more personal finance context. You can connect up to 30 accounts, up from 5, and Computer now fetches up to 500 assets per account, up from 200.", "artifact_url": "https://www.perplexity.ai/changelog/personal-computer-for-all-users-on-mac", "our_archive": "ledger/archives/2026-05/perplexity/e338a812-www-perplexity-ai-changelog-personal-computer-for-all-users-on-mac/20260725T151659Z", "bundle_sha256": "804bbf73b1770c727d74e4de222c2798097a183f4130b1dae8f5344583b11942", "source": "reported" }, { "id": "E-2026-056", "date": "2026-05-07", "vendor": "Cursor", "headline": "Plan execution gained a parallel mode using async subagents", "what_plain": "Cursor's changelog states that plan execution changed from doing tasks one at a time to multitasking across tasks (a new 'Build in Parallel' mode using async subagents), stating both the old and new mechanism in the same sentence.", "quote": "Cursor can now execute on plans faster by multitasking across tasks instead of tackling them one at a time. Click \"Build in Parallel\" to have it identify independent parts of your plan and run them simultaneously using async subagents. Cursor will keep dependent steps in order when needed.", "artifact_url": "https://cursor.com/changelog/05-07-26", "our_archive": "ledger/archives/2026-05/cursor/29f64665-cursor-com-changelog-05-07-26/20260725T152803Z", "bundle_sha256": "5bb544407ddb935ab26b1cbe3d70c9eadf92ac9c20a4802148ebdaeae9b02b1b", "source": "reported" }, { "id": "E-2026-042", "date": "2026-05-06", "vendor": "Anthropic", "headline": "Claude Code's five hour rate limits were doubled for Pro, Max, Team and Enterprise plans", "what_plain": "Anthropic's own announcement of a compute partnership with SpaceX also states three concurrent, dated, 'effective today' changes to Claude usage limits: Claude Code's five-hour rate limits were doubled for Pro/Max/Team/seat-based Enterprise plans; a previously-existing peak-hours limit reduction on Claude Code was removed for Pro/Max accounts; and API rate limits for Claude Opus models were raised, shown as an explicit before-to-after numeric table (e.g. Tier 1: 30,000 to 500,000 max input tokens/minute, 8,000 to 80,000 max output tokens/minute). A previously unrecorded same-label finding -- not the reasoning-effort story covered by E-2026-038/039/040 or the postmortem regressions in E-2026-TBD-1/2, and not surfaced by the original bounded pass (which did not read this page).", "quote": "First, we're doubling Claude Code’s five-hour rate limits for Pro, Max, Team, and seat-based Enterprise plans. Second, we're removing the peak hours limit reduction on Claude Code for Pro and Max accounts. Third, we're raising our API rate limits considerably for Claude Opus models, as shown in the table below: [table caption: 'Our updated API rate limits for Claude Opus models.']", "artifact_url": "https://www.anthropic.com/news/higher-limits-spacex", "our_archive": "ledger/archives/2026-04/anthropic/f53c308c-www-anthropic-com-news-higher-limits-spacex/20260725T173006Z", "bundle_sha256": "07df8d529d8741f9d6c8191032a9509da1790e37822643d28eff953ece5203b3", "source": "reported", "date_note": "2026-05-06 (in grace window)" }, { "id": "E-2026-057", "date": "2026-05-06", "vendor": "Google", "headline": "The Interactions API replaced its outputs array response schema with a steps array", "what_plain": "Google's Gemini API changelog and its linked migration guide describe the Interactions API's response schema changing from a flat `outputs` array to a `steps` array (plus `response_format` output-config changes) on the same API endpoint, with the new schema becoming the default on May 26, 2026 and the legacy schema being permanently removed on June 8, 2026. The legacy state, the new state, and both effective dates are all stated in the same first-party artifact.", "quote": "Upcoming breaking change: The Interactions API request and response schema (outputs → steps) and output format configuration (response_format) are changing. The new schema becomes the default on May 26 and the legacy schema will be removed on June 8. -- Core change: outputs to steps. The new schema replaces the outputs array with a steps array. Legacy: Responses returned a flat outputs array containing only the model's generated content. New schema: Responses return a steps array containing structured steps with type discriminators. -- After May 26, the new schema becomes the default for all requests. You can temporarily opt out with Api-Revision: 2026-05-07 until June 8, when the API permanently removes the legacy schema.", "artifact_url": "https://ai.google.dev/gemini-api/docs/interactions-breaking-changes-may-2026", "our_archive": "ledger/archives/2026-05/google/bad2e3e4-ai-google-dev-gemini-api-docs-interactions-breaking-changes-may-2026/20260725T165435Z", "bundle_sha256": "a15da68e136e84b8aeee5ed09176d877cedc808e65603b1727c565a4a5cf38ae", "source": "reported", "date_note": "2026-05-06 (changelog disclosure; effective dates in the same artifact: May 26, 2026 default flip, June 8, 2026 legacy removal)" }, { "id": "E-2026-044", "date": "2026-05-05", "vendor": "Microsoft", "headline": "Viva Insights Power BI reports could now be published to anyone in the organization", "what_plain": "Microsoft's Microsoft 365 Copilot release notes state that Viva Insights Power BI report publishing, previously limited to recipients who held Copilot Dashboard access, was changed as of this dated entry to allow publishing to any user in the organization.", "quote": "What changed: Previously, analysts could publish Viva Insights Power BI reports only to people who had access to the Copilot Dashboard. With this update, analysts can publish these reports to anyone in their organization. This broadens visibility and removes the dependency on global Copilot Dashboard access to receive reports.", "artifact_url": "https://learn.microsoft.com/en-us/microsoft-365/copilot/release-notes", "our_archive": "ledger/archives/2026-04/microsoft/7856f251-learn-microsoft-com-en-us-microsoft-365-copilot-release-notes/20260725T151304Z", "bundle_sha256": "b448a11e8ac3f6d38e5f57b540dd5fac9174bda7dc8514f5a450bd7c4f0e703b", "source": "reported" }, { "id": "E-2026-046", "date": "2026-05-05", "vendor": "OpenAI", "headline": "GPT-5.5 Instant replaced GPT-5.3 Instant as ChatGPT's default model", "what_plain": "OpenAI made GPT-5.5 Instant the new default model in ChatGPT, replacing GPT-5.3 Instant for all ChatGPT users; GPT-5.3 Instant remains available to paid users for three months through model configuration settings before it is retired.", "quote": "We’re rolling out GPT-5.5 Instant as ChatGPT’s new default model. It replaces GPT-5.3 Instant for all ChatGPT users and improves everyday answers across accuracy, clarity and conciseness, image understanding, STEM questions, and when ChatGPT decides to search the web. [...] GPT-5.5 Instant is rolling out starting today to all ChatGPT users. Paid users can continue using GPT-5.3 Instant for three months through model configuration settings before it is retired.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-04/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151551Z", "bundle_sha256": "45a6c5bb090bf6e42d4afde9121e3f3ea455ba33824d9cf7b0dd91ab2ef87773", "source": "reported" }, { "id": "E-2026-043", "date": "2026-04-24", "vendor": "DeepSeek", "headline": "The deepseek-chat and deepseek-reasoner aliases now point to deepseek-v4-flash", "what_plain": "DeepSeek's own API Change Log page shows, in one continuous dated list, that the identifiers deepseek-chat and deepseek-reasoner pointed to DeepSeek-V3.2 as of the 2025-12-01 entry, and to deepseek-v4-flash (non-thinking and thinking modes respectively) as of the 2026-04-24 entry, ahead of the two legacy names' stated discontinuation on 2026-07-24.", "quote": "Date: 2026-04-24 -- DeepSeek-V4: \"The two legacy API model names, deepseek-chat and deepseek-reasoner, will be discontinued in three months (2026-07-24). During the current period, these two model names point to the non-thinking mode and thinking mode of deepseek-v4-flash, respectively.\" Immediately preceding entry on the same page, Date: 2025-12-01 -- DeepSeek-V3.2: \"Both deepseek-chat and deepseek-reasoner have been upgraded to DeepSeek-V3.2. deepseek-chat corresponds to DeepSeek-V3.2's non-thinking mode. deepseek-reasoner corresponds to DeepSeek-V3.2's thinking mode.\"", "artifact_url": "https://api-docs.deepseek.com/updates", "our_archive": "ledger/archives/2026-04/deepseek/8f8555b7-api-docs-deepseek-com-updates/20260725T151047Z", "bundle_sha256": "7912c5fd081c82f3d638b71d5adc241724f2b650b22b9a72729d0f0f54a622e9", "source": "reported" }, { "id": "E-2026-039", "date": "2026-04-22", "vendor": "Anthropic", "headline": "Claude Code's default effort level for Opus 4.6 and Sonnet 4.6 rose from medium to high", "what_plain": "Anthropic's Claude Code changelog recorded, in the entry for release v2.1.117, that the default effort level for Pro/Max subscribers on Opus 4.6 and Sonnet 4.6 was now high, stating it had previously been medium.", "quote": "Default effort for Pro/Max subscribers on Opus 4.6 and Sonnet 4.6 is now `high` (was `medium`)", "artifact_url": "https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md", "our_archive": "ledger/archives/2026-04/anthropic/79752d66-github-com-anthropics-claude-code-blob-main-changelog-md/20260725T161114Z", "bundle_sha256": "c7ae3c388457c01b013038c0ceb53ebe42adea4704ea86802f7986777f172037", "source": "reported" }, { "id": "E-2026-054", "date": "2026-04-18", "vendor": "xAI", "headline": "Imagine began filtering moderated items without media out of its lists", "what_plain": "Grok's own consumer release notes stated, under an April 18 2026 bug-fix entry, that moderated items without media are now filtered out of Imagine's lists.", "quote": "Moderated items without media are now filtered out of lists.", "artifact_url": "https://grok.com/release-notes", "our_archive": "ledger/archives/2026-04/xai/041f1d41-grok-com-release-notes/20260725T151258Z", "bundle_sha256": "174ef514caa805996493fc92a6db70e46988faad3452ef2ad604a3e21170926b", "source": "reported" }, { "id": "E-2026-040", "date": "2026-04-16", "vendor": "Anthropic", "headline": "Claude Code's default effort level was raised to xhigh for all plans", "what_plain": "In its announcement of Claude Opus 4.7, Anthropic stated that in Claude Code it had raised the default effort level to xhigh for all plans, alongside introducing the new xhigh effort level.", "quote": "In Claude Code, we've raised the default effort level to xhigh for all plans.", "artifact_url": "https://www.anthropic.com/news/claude-opus-4-7", "our_archive": "ledger/archives/2026-04/anthropic/b8af8104-www-anthropic-com-news-claude-opus-4-7/20260725T150716Z", "bundle_sha256": "8d366d38267c028a5f6291dd75b80748e95f0c79b3466c255277fee39b11b617", "source": "reported" }, { "id": "E-2026-041", "date": "2026-04-16", "vendor": "Anthropic", "headline": "A verbosity cutting system prompt hurt Claude's coding quality alongside other changes", "what_plain": "The same Anthropic postmortem states that on April 16 it added a system prompt instruction to reduce verbosity, which, in combination with other prompt changes, hurt coding quality for Sonnet 4.6, Opus 4.6, and Opus 4.7, and was reverted on April 20. A second, distinct same-label finding from the same artifact as E-2026-TBD-1, previously unrecorded and not surfaced by the original bounded pass.", "quote": "On April 16, we added a system prompt instruction to reduce verbosity. In combination with other prompt changes, it hurt coding quality and was reverted on April 20. This impacted Sonnet 4.6, Opus 4.6, and Opus 4.7.", "artifact_url": "https://www.anthropic.com/engineering/april-23-postmortem", "our_archive": "ledger/archives/2026-04/anthropic/85b0ca3c-www-anthropic-com-engineering-april-23-postmortem/20260725T170027Z", "bundle_sha256": "23fb916554a2b29fa9960c8a3093ca2c89686c027ef52f15243994045891f98f", "source": "reported", "date_note": "2026-04-16 (shipped) / 2026-04-20 (reverted) / 2026-04-23 (disclosed, in-window)" }, { "id": "E-2026-050", "date": "2026-04-16", "vendor": "Perplexity", "headline": "Claude Opus 4.7 became the default orchestrator model for Computer", "what_plain": "Perplexity's own changelog states, in a 2026-04-16 dated post, that Claude Opus 4.7 became the default orchestrator model for its 'Computer' agent product; a 2026-05-03 dated post in the same changelog series states that GPT-5.5 then became the new default orchestration model for the same product.", "quote": "Claude Opus 4.7 is now the default orchestrator model for Computer. Choose Claude Opus 4.7 for higher-quality reasoning, writing, and coding tasks. As of today, you can now choose between Claude Opus 4.7, GPT-5.4, and Claude Sonnet 4.6 to be the main orchestrator model.", "artifact_url": "https://www.perplexity.ai/changelog/personal-computer-on-mac-launch-and-computer-updates---april-17-2026", "our_archive": "ledger/archives/2026-04/perplexity/d5913f1e-www-perplexity-ai-changelog-personal-computer-on-mac-launch-and-computer-updates/20260725T151503Z", "bundle_sha256": "655aef16e1c9455bd0cf41b03b97935cc3be51e5fb6ff31bc77ddc10c979fcd8", "source": "reported" }, { "id": "E-2026-052", "date": "2026-04-11", "vendor": "xAI", "headline": "Grok restored the bot challenge check on file upload", "what_plain": "Grok's own consumer release notes stated, under an April 11 2026 bug-fix entry, that the bot-challenge check on file upload was restored.", "quote": "Restored the bot-challenge check on file upload.", "artifact_url": "https://grok.com/release-notes", "our_archive": "ledger/archives/2026-04/xai/041f1d41-grok-com-release-notes/20260725T151258Z", "bundle_sha256": "174ef514caa805996493fc92a6db70e46988faad3452ef2ad604a3e21170926b", "source": "reported" }, { "id": "E-2026-053", "date": "2026-04-11", "vendor": "xAI", "headline": "Imagine's NSFW content setting toggle was fixed", "what_plain": "Grok's own consumer release notes stated, in the same April 11 2026 bug-fix entry, that the Imagine NSFW content-setting toggle was fixed and a check was added to video-generation presets.", "quote": "Fixed the NSFW setting toggle and added a check on video generation presets.", "artifact_url": "https://grok.com/release-notes", "our_archive": "ledger/archives/2026-04/xai/041f1d41-grok-com-release-notes/20260725T151258Z", "bundle_sha256": "174ef514caa805996493fc92a6db70e46988faad3452ef2ad604a3e21170926b", "source": "reported" }, { "id": "E-2026-047", "date": "2026-04-09", "vendor": "OpenAI", "headline": "Codex usage on ChatGPT Plus shifted from long single day sessions to spread out sessions", "what_plain": "OpenAI changed how Codex usage works on the existing ChatGPT Plus plan: a temporary Codex usage promotion on Plus ended, and Plus usage was rebalanced away from longer high-intensity single-day Codex sessions toward more sessions spread across the week -- same plan, same price, different usage behaviour, both states stated in OpenAI's own announcement.", "quote": "We’re introducing a new $100/month Pro plan and updating how Codex usage works across Plus and Pro. [...] Plus ($20/month): Plus remains best for steady, day-to-day use. As the temporary Codex promotion on Plus ends, we’re rebalancing Plus usage to support more sessions across the week, rather than longer, high-intensity sessions on a single day.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-04/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151626Z", "bundle_sha256": "dd0fec1e90e4dd3c5ef7bb886e374f1c3170156e768fa82e1f82fbfd5117766f", "source": "reported" }, { "id": "E-2026-045", "date": "2026-04-08", "vendor": "Microsoft", "headline": "Copilot's codebase search tool switched to a single auto managed semantic index", "what_plain": "GitHub's official Changelog post for GitHub Copilot in VS Code's March 2026 releases states that the #codebase tool changed from supporting local-versus-remote indexing to running purely semantic searches against a single auto-managed index.", "quote": "The #codebase tool now runs purely semantic searches against a single auto-managed index. No more local vs. remote indexing—just faster, more consistent results.", "artifact_url": "https://github.blog/changelog/2026-04-08-github-copilot-in-visual-studio-code-march-releases/", "our_archive": "ledger/archives/2026-04/microsoft/83d010f9-github-blog-changelog-2026-04-08-github-copilot-in-visual-studio-code-march-rele/20260725T151949Z", "bundle_sha256": "ffb9545c6ecc415ae123ac8e6957c1f92671a6ce200cf36bbb5916b361f0a854", "source": "reported" }, { "id": "E-2026-038", "date": "2026-04-07", "vendor": "Anthropic", "headline": "Claude Code's default effort level rose from medium to high for API and enterprise users", "what_plain": "Anthropic's Claude Code changelog recorded, in the entry for release v2.1.94, that the default effort level changed from medium to high for API-key, Bedrock/Vertex/Foundry, Team, and Enterprise users.", "quote": "Changed default effort level from medium to high for API-key, Bedrock/Vertex/Foundry, Team, and Enterprise users (control this with `/effort`)", "artifact_url": "https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md", "our_archive": "ledger/archives/2026-04/anthropic/79752d66-github-com-anthropics-claude-code-blob-main-changelog-md/20260725T161006Z", "bundle_sha256": "12783af4d947bd7cde0ce612eca92062616589f54765289ed8332f635d775443", "source": "reported" }, { "id": "E-2026-049", "date": "2026-04-07", "vendor": "OpenAI", "headline": "Codex stopped showing several legacy models in its picker for ChatGPT sign in users", "what_plain": "OpenAI's own Codex changelog (learn.chatgpt.com/docs/changelog, the URL help.openai.com/en/articles/11428266-codex-changelog redirects to) -- a distinct, previously-unread surface from the github.com/openai/codex CLI releases the original chase checked -- states that starting April 7, 2026, the model picker for users who sign in to Codex with ChatGPT stopped showing six named models (gpt-5.2-codex, gpt-5.1-codex-mini, gpt-5.1-codex-max, gpt-5.1-codex, gpt-5.1, gpt-5), with full removal from Codex on April 14. This directly touches the Codex bucket's original allegation cluster (a report that 'Codex started blocking tool calls,' GitHub issue #20364) without being the same claim -- it is a genuine, dated, explicit model-availability contrast the original bounded pass did not have, because it never read this changelog surface (learn.chatgpt.com/docs/changelog is a separate site from github.com/openai/codex, and the original chase's Codex coverage was limited to GitHub CLI releases).", "quote": "We're updating model availability for users who sign in with ChatGPT. Starting April 7, the model picker no longer shows gpt-5.2-codex, gpt-5.1-codex-mini, gpt-5.1-codex-max, gpt-5.1-codex, gpt-5.1, or gpt-5. On April 14, we'll remove those models from Codex for ChatGPT sign-in. Users can still choose from gpt-5.4, gpt-5.4-mini, gpt-5.3-codex, and gpt-5.2. ChatGPT Pro users can also choose gpt-5.3-codex-spark.", "artifact_url": "https://help.openai.com/en/articles/11428266-codex-changelog (redirects to https://learn.chatgpt.com/docs/changelog)", "our_archive": "ledger/archives/2026-04/openai/a77e6b1d-learn-chatgpt-com-docs-changelog/20260725T171544Z", "bundle_sha256": "c255342a533be481a54db7aeaeb29ca5b80a65017970129667b6b65d7c485bc2", "source": "reported" }, { "id": "E-2026-051", "date": "2026-04-04", "vendor": "xAI", "headline": "Imagine's video extension start time was clamped to a minimum of two seconds", "what_plain": "Grok's own consumer release notes stated, under an April 4 2026 bug-fix entry, that Imagine's video-extension start time is now clamped to a minimum of two seconds.", "quote": "Video extension start time is now clamped to a minimum of two seconds.", "artifact_url": "https://grok.com/release-notes", "our_archive": "ledger/archives/2026-04/xai/041f1d41-grok-com-release-notes/20260725T151258Z", "bundle_sha256": "174ef514caa805996493fc92a6db70e46988faad3452ef2ad604a3e21170926b", "source": "reported" }, { "id": "E-2026-026", "date": "2026-03-26", "vendor": "Anthropic", "headline": "A bug made Claude clear its recent thinking every turn instead of after idle sessions", "what_plain": "The same April 23 postmortem disclosed a bug, shipped March 26, that caused Claude to repeatedly clear its own recent thinking every turn instead of only after idle sessions, making it seem forgetful and repetitive; fixed April 10; affected Sonnet 4.6 and Opus 4.6.", "quote": "On March 26, we shipped a change to clear Claude's older thinking from sessions that had been idle for over an hour, to reduce latency when users resumed those sessions. A bug caused this to keep happening every turn for the rest of the session instead of just once, which made Claude seem forgetful and repetitive. We fixed it on April 10. This affected Sonnet 4.6 and Opus 4.6.", "artifact_url": "https://www.anthropic.com/engineering/april-23-postmortem", "our_archive": "ledger/archives/2026-03/anthropic/85b0ca3c-www-anthropic-com-engineering-april-23-postmortem/20260725T150437Z", "bundle_sha256": "c57907569cad2b28d7de78ce0ed4be831bf6e0a6a8451e9b291c7d6eb40ecb74", "source": "reported", "date_note": "2026-03-26 (disclosed 2026-04-23)" }, { "id": "E-2026-037", "date": "2026-03-26", "vendor": "Perplexity", "headline": "Computer began drafting output in Markdown by default instead of PDF", "what_plain": "Perplexity's own changelog disclosed that Computer now drafts output in lightweight Markdown by default instead of generating PDFs, with PDF/DOCX still available as an on-demand export.", "quote": "Markdown-first drafting with export on demand: Computer now drafts in lightweight Markdown by default instead of generating PDFs. Markdown outputs appear as clean reports in the thread, and can also be downloaded as PDF or DOCX at any time.", "artifact_url": "https://www.perplexity.ai/changelog/what-we-shipped--march-27-2026", "our_archive": "ledger/archives/2026-03/perplexity/97664583-www-perplexity-ai-changelog-what-we-shipped-march-27-2026/20260725T152122Z", "bundle_sha256": "1b144ee062c023d7819268f0e2636639787e2cabfbaad1b3736503b4e5901657", "source": "reported", "date_note": "2026-03-26 (published in the March 27, 2026 weekly digest)" }, { "id": "E-2026-034", "date": "2026-03-17", "vendor": "OpenAI", "headline": "ChatGPT retired its Nerdy personality preset and moved users to the default", "what_plain": "OpenAI disclosed that it was sunsetting the preset 'Nerdy' personality/base style in ChatGPT; users who had it selected were moved to the default personality.", "quote": "Sunsetting the “Nerdy” base style. We’re sunsetting the preset Nerdy base style to simplify options and improve response quality. Users who selected it will be moved to the default personality. You can manage personality settings anytime in ChatGPT under Settings → Personalization.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-03/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151341Z", "bundle_sha256": "7c262b1f3c47cd40741c14df1ceb360229ef6683d5839022441ec18b6b8479d6", "source": "reported" }, { "id": "E-2026-033", "date": "2026-03-16", "vendor": "OpenAI", "headline": "GPT-5.3 Instant reduced teaser style phrasing patterns in its replies", "what_plain": "A second, separate GPT-5.3 Instant tone update, disclosed two weeks after E-2026-032, aimed at follow-up tone and reducing 'teaser-style' phrasing patterns like 'If you want…' and 'You'll never believe…'.", "quote": "GPT-5.3 Instant update. We’re rolling out an update to GPT-5.3 Instant that improves follow-up tone and reduces teaser-style phrasing in responses (e.g., “If you want…”, “You’ll never believe…”, “I can tell you these three things that…”).", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-03/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151341Z", "bundle_sha256": "7c262b1f3c47cd40741c14df1ceb360229ef6683d5839022441ec18b6b8479d6", "source": "reported" }, { "id": "E-2026-035", "date": "2026-03-16", "vendor": "OpenAI", "headline": "The gpt-5.3-chat-latest alias now points to whatever model ChatGPT currently uses", "what_plain": "OpenAI's API changelog disclosed that the gpt-5.3-chat-latest alias was repointed to whatever model ChatGPT currently uses -- the same public slug now serving a different underlying model, by design.", "quote": "Mar 16 Update gpt-5.3-chat-latest. Updated the gpt-5.3-chat-latest slug to point to the latest model currently used in ChatGPT.", "artifact_url": "https://developers.openai.com/api/docs/changelog", "our_archive": "ledger/archives/2026-03/openai/49e8d057-developers-openai-com-api-docs-changelog/20260725T151409Z", "bundle_sha256": "f924c8ebd7d880824419bf2a357fb89183dfacb5ceee90628a1075c51f5790a7", "source": "reported" }, { "id": "E-2026-030", "date": "2026-03-13", "vendor": "Anthropic", "headline": "The 1M token context window became generally available for Opus 4.6 and Sonnet 4.6", "what_plain": "Anthropic's Claude API release notes disclosed that the 1M token context window became generally available for Opus 4.6 and Sonnet 4.6 at standard pricing (previously beta), that dedicated 1M rate limits were removed for all supported models, and that the media limit for 1M-context requests was raised from 100 to 600 images/PDF pages.", "quote": "The 1M token context window is now generally available for Claude Opus 4.6 and Sonnet 4.6 at standard pricing. Requests over 200k tokens work automatically for these models with no beta header required. The 1M token context window remains in beta for Claude Sonnet 4.5 and Sonnet 4. We've removed the dedicated 1M rate limits for all supported models. Your standard account limits now apply across every context length. We've raised the media limit from 100 to 600 images or PDF pages per request when using the 1M token context window.", "artifact_url": "https://platform.claude.com/docs/en/release-notes/overview", "our_archive": "ledger/archives/2026-03/anthropic/4182359e-platform-claude-com-docs-en-release-notes-overview/20260725T150822Z", "bundle_sha256": "ca572f0bcfe2a78720d028755f4fb53514fc6dcf66a21235d4277450bdc3c231", "source": "reported" }, { "id": "E-2026-036", "date": "2026-03-13", "vendor": "OpenAI", "headline": "GPT-5.4 got an image encoder fix for a bug in input image processing", "what_plain": "OpenAI's API changelog disclosed a fix to its image encoder for a bug affecting input_image processing in GPT-5.4, stating some image-understanding use cases may now see improved quality with no action required from callers.", "quote": "Mar 13 Fix gpt-5.4 v1/responses v1/chat/completions. Updated our image encoder to fix a small bug with input_image inputs in GPT-5.4. Some image understanding use cases may now see improved quality. No action is required.", "artifact_url": "https://developers.openai.com/api/docs/changelog", "our_archive": "ledger/archives/2026-03/openai/49e8d057-developers-openai-com-api-docs-changelog/20260725T151409Z", "bundle_sha256": "f924c8ebd7d880824419bf2a357fb89183dfacb5ceee90628a1075c51f5790a7", "source": "reported" }, { "id": "E-2026-028", "date": "2026-03-12", "vendor": "Anthropic", "headline": "Claude Code's default Opus model on Bedrock, Vertex and Foundry moved from 4.1 to 4.6", "what_plain": "Claude Code's release notes for v2.1.73 disclosed that the default Opus model on Bedrock, Vertex, and Microsoft Foundry changed from Opus 4.1 to Opus 4.6.", "quote": "Changed default Opus model on Bedrock, Vertex, and Microsoft Foundry to Opus 4.6 (was Opus 4.1).", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.73", "our_archive": "ledger/archives/2026-03/anthropic/5e4d08c1-github-com-anthropics-claude-code-releases-tag-v2-1-73/20260725T150555Z", "bundle_sha256": "069fb1924c2767161a0235fe7c9f0ce9194a904c2dab58ae8d82f74565c19f64", "source": "reported" }, { "id": "E-2026-031", "date": "2026-03-09", "vendor": "Google", "headline": "The gemini-3-pro-preview alias now points to a different model", "what_plain": "Google's Gemini API changelog disclosed that the Gemini 3 Pro Preview model was shut down and that the gemini-3-pro-preview alias now points to gemini-3.1-pro-preview -- the same public slug now serving a different model.", "quote": "March 9, 2026. The Gemini 3 Pro Preview model has been shut down. The gemini-3-pro-preview now points to gemini-3.1-pro-preview.", "artifact_url": "https://ai.google.dev/gemini-api/docs/changelog", "our_archive": "ledger/archives/2026-03/google/a24fd97e-ai-google-dev-gemini-api-docs-changelog/20260725T151643Z", "bundle_sha256": "963fe4ccbaf03ae39662ccdc397745cfb178d2dec533a6f898e5c471514a4c6a", "source": "reported" }, { "id": "E-2026-027", "date": "2026-03-05", "vendor": "Anthropic", "headline": "Claude Code moved Sonnet 4.5 users on Pro, Max and Team Premium plans to Sonnet 4.6", "what_plain": "Claude Code's release notes for v2.1.69 disclosed that Sonnet 4.5 users on Pro/Max/Team Premium plans were automatically migrated to Sonnet 4.6.", "quote": "Changed Sonnet 4.5 users on Pro/Max/Team Premium to be automatically migrated to Sonnet 4.6.", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.69", "our_archive": "ledger/archives/2026-03/anthropic/bd08d372-github-com-anthropics-claude-code-releases-tag-v2-1-69/20260725T150625Z", "bundle_sha256": "8eec94fcdd0274bf6f7f1185d6fe0dc31063203f2cee937d7d2692b723945537", "source": "reported" }, { "id": "E-2026-024", "date": "2026-03-04", "vendor": "Anthropic", "headline": "Claude Code removed Opus 4 and 4.1, moving pinned users to Opus 4.6", "what_plain": "Claude Code's own release notes for v2.1.68 (published same-day) disclosed, without further comment, that Opus 4.6 now defaults to medium effort for Max and Team subscribers, and that Opus 4 and 4.1 were removed from the first-party API with pinned users auto-moved to Opus 4.6.", "quote": "Opus 4.6 now defaults to medium effort for Max and Team subscribers. Medium effort works well for most tasks — it's the sweet spot between speed and thoroughness. You can change this anytime with `/model`. Re-introduced the \"ultrathink\" keyword to enable high effort for the next turn. Removed Opus 4 and 4.1 from Claude Code on the first-party API — users with these models pinned are automatically moved to Opus 4.6.", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.68", "our_archive": "ledger/archives/2026-03/anthropic/9fc08856-github-com-anthropics-claude-code-releases-tag-v2-1-68/20260725T150740Z", "bundle_sha256": "7e6615ffa7325085e4df80ba6e768db62c50ff43b1e7e477ae3474b002448557", "source": "reported" }, { "id": "E-2026-025", "date": "2026-03-04", "vendor": "Anthropic", "headline": "Claude Code's default effort level was reverted from medium back to high", "what_plain": "Anthropic's engineering postmortem, published April 23, stated that on March 4 it changed Claude Code's default reasoning effort from high to medium to reduce latency, called this the wrong tradeoff, and said the change was reverted on April 7 -- affecting both Sonnet 4.6 and Opus 4.6.", "quote": "On March 4, we changed Claude Code's default reasoning effort from high to medium to reduce the very long latency—enough to make the UI appear frozen—some users were seeing in high mode. This was the wrong tradeoff. We reverted this change on April 7 after users told us they'd prefer to default to higher intelligence and opt into lower effort for simple tasks. This impacted Sonnet 4.6 and Opus 4.6.", "artifact_url": "https://www.anthropic.com/engineering/april-23-postmortem", "our_archive": "ledger/archives/2026-03/anthropic/85b0ca3c-www-anthropic-com-engineering-april-23-postmortem/20260725T150437Z", "bundle_sha256": "c57907569cad2b28d7de78ce0ed4be831bf6e0a6a8451e9b291c7d6eb40ecb74", "source": "reported", "date_note": "2026-03-04 (fuller disclosure 2026-04-23)" }, { "id": "E-2026-032", "date": "2026-03-03", "vendor": "OpenAI", "headline": "GPT-5.3 Instant began avoiding dead ends, caveats and overly declarative phrasing", "what_plain": "OpenAI's ChatGPT release notes disclosed an update to GPT-5.3 Instant intended to reduce unnecessary dead ends, caveats, and overly declarative phrasing, aimed at tone, relevance, and conversational flow.", "quote": "GPT-5.3 Instant Update. GPT‑5.3 Instant delivers more accurate answers, richer and better-contextualized results when searching the web, and reduces unnecessary dead ends, caveats, and overly declarative phrasing that can interrupt the flow of conversation. This update focuses on the parts of the ChatGPT experience people feel every day: tone, relevance, and conversational flow. These are nuanced problems that don’t always show up in benchmarks, but shape whether ChatGPT feels helpful or frustrating. GPT‑5.3 Instant directly reflects user feedback in these areas.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-03/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151341Z", "bundle_sha256": "7c262b1f3c47cd40741c14df1ceb360229ef6683d5839022441ec18b6b8479d6", "source": "reported" }, { "id": "E-2026-014", "date": "2026-02-24", "vendor": "Anthropic", "headline": "Claude Code's Bash tool began skipping the login shell by default", "what_plain": "Three explicit default-behaviour changes logged on the same dated Claude Code release: the Bash tool's login-shell flag default flipped (previously required an env var to skip it, now skips it by default when a snapshot is available), the tool-result disk-persistence threshold was halved, and the plugin marketplace's git timeout was quadrupled.", "quote": "BashTool now skips login shell (-l flag) by default when a shell snapshot is available, improving command execution performance. Previously this required setting CLAUDE_BASH_NO_LOGIN=true. | Tool results larger than 50K characters are now persisted to disk (previously 100K). This reduces context window usage and improves conversation longevity. | Updated plugin marketplace default git timeout from 30s to 120s and added CLAUDE_CODE_PLUGIN_GIT_TIMEOUT_MS to configure.", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.51", "our_archive": "ledger/archives/2026-02/anthropic/ebf7c4c6-github-com-anthropics-claude-code-releases-tag-v2-1-51/20260725T151451Z", "bundle_sha256": "27111532df9a17d1c5b08a8b11516276bebbfb322f656241e00e10cdb5081423", "source": "reported" }, { "id": "E-2026-020", "date": "2026-02-20", "vendor": "OpenAI", "headline": "ChatGPT's Thinking mode context window grew from 196k to 256k tokens", "what_plain": "OpenAI's own ChatGPT release notes state the total context window for manually-selected Thinking mode expanded from 196k to 256k tokens (128k input, 128k max output) -- an explicit numeric before/after for the same named mode.", "quote": "February 20, 2026. Expanded context window for Thinking. ChatGPT now has a total context window length of 256k tokens (128k input and 128k max output) when you manually select Thinking. Previously it was 196k total tokens.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-02/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T151909Z", "bundle_sha256": "318697a2264cd0a1dbd32d2c7d2efe70b6d592d561417889b78767075f587e9e", "source": "reported" }, { "id": "E-2026-013", "date": "2026-02-19", "vendor": "Anthropic", "headline": "Claude Code moved the Max plan's 1M context from Sonnet 4.5 to Sonnet 4.6", "what_plain": "Claude Code's own release notes state that Sonnet 4.5 with 1M context was being removed from the Max plan, with users directed to switch to Sonnet 4.6 which now carries the 1M context capability instead.", "quote": "Sonnet 4.5 with 1M context is being removed from the Max plan in favor of our frontier Sonnet 4.6 model, which now has 1M context. Please switch in /model.", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.49", "our_archive": "ledger/archives/2026-02/anthropic/40a2608a-github-com-anthropics-claude-code-releases-tag-v2-1-49/20260725T151414Z", "bundle_sha256": "37b69050c6bc5d3b9e7f6e90cb7ba6867f3845561de4e6102d7cab00875e2022", "source": "reported" }, { "id": "E-2026-017", "date": "2026-02-17", "vendor": "Anthropic", "headline": "Sonnet 4.6's BrowseComp scores were revised down after a cheating detection fix", "what_plain": "The Claude Sonnet 4.6 System Card's own internal Changelog section states the same class of correction as E-2026-016: an improved cheating-detection pipeline flagged additional unintended solutions, and both BrowseComp scores were revised down. The system card itself is dated February 17, 2026; the correction was added to its changelog on March 6, 2026. Anthropic's engineering blog post the same day states it updated the model cards for both Opus 4.6 and Sonnet 4.6, but publishes no Sonnet figures -- Sonnet's corrected numbers appear only in this system card.", "quote": "Mar 6, 2026: Sonnet 4.6's BrowseComp scores updated due to running an improved cheating detection pipeline. Section 2.20.1.1 and fig. 2.20.1.1.A: Updated highest single-agent BrowseComp score (74.72% -> 74.01%). Improved pipeline flagged 9 additional instances of unintended solutions to the problems that our original pipeline had missed. Instead of re-running the evaluation, we marked those 9 solutions as incorrect. Only the best-scoring single-agent config was re-verified. | Section 2.20.1.3: Updated multi-agent BrowseComp (82.62% -> 82.07%). Improved pipeline flagged 11 instances of unintended solutions to the problems. After updating the blacklist and re-running the problems with the leaks removed, 5/11 problems remained correct. | Corroborating, from the engineering blog: \"We have updated the model cards for both Claude Opus 4.6 and Claude Sonnet 4.6.\"", "artifact_url": "https://www.anthropic.com/claude-sonnet-4-6-system-card", "our_archive": "ledger/archives/2026-02/anthropic/8e91d5e0-www-cdn-anthropic-com-bbd8ef16d70b7a1665f14f306ee88b53f686aa75-claude-20sonnet-2/20260725T152110Z", "bundle_sha256": "c7cdb4e23d0f2d5550764f9f1c3a2ce4c03ee3add1fdc64ff1397c98c58a67c0", "source": "reported", "date_note": "2026-02-17 (system card published) / 2026-03-06 (changelog correction, within February's 7-day grace)" }, { "id": "E-2026-016", "date": "2026-02-10", "vendor": "Anthropic", "headline": "Opus 4.6's benchmark scores were revised down after a cheating detection pipeline update", "what_plain": "The Claude Opus 4.6 System Card's own internal Changelog section states four corrections to previously-published benchmark scores for Opus 4.6 -- an improved cheating-detection pipeline flagged additional instances of solutions it should have caught, and the published numbers were revised down.", "quote": "Feb 10, 2026: Fixed reported score in section 2.19.2 for MMMU-Pro (\"70.7%\" -> \"70.6%\"). | Feb 17, 2026: Updated reported score for Opus 4.6 in section 2.21.2 for HLE with tools (53.1% -> 53.0%). The update was caused by running an improved cheating detection pipeline which flagged 3 additional instances of cheating that our original pipeline had missed. | Mar 6, 2026: Opus 4.6's BrowseComp scores updated due to running an improved cheating detection pipeline. Updated highest single-agent BrowseComp score (83.97% -> 83.73%). ... Updated multi-agent BrowseComp (86.81% -> 86.57%).", "artifact_url": "https://www.anthropic.com/claude-opus-4-6-system-card", "our_archive": "ledger/archives/2026-02/anthropic/9eff230a-www-cdn-anthropic-com-6a5fa276ac68b9aeb0c8b6af5fa36326e0e166dd-claude-20opus-204/20260725T154146Z", "bundle_sha256": "90287dfa7ebff924167c8f67a9b5352fea474355f52b56909d4ebe130ae6bc71", "source": "reported", "date_note": "2026-02-10 / 2026-02-17 / 2026-03-06 (grace)" }, { "id": "E-2026-021", "date": "2026-02-05", "vendor": "Perplexity", "headline": "Deep Research began running on Opus 4.6 instead of Opus 4.5", "what_plain": "Perplexity's own changelog states Deep Research ran on Opus 4.5 as of the Feb 5/6 entry, then states one week later that Deep Research runs on Opus 4.6 -- the same product feature (Deep Research) served by a different underlying model, both states in Perplexity's own dated changelog entries.", "quote": "Feb 5/6, 2026: ...with Deep Research now running on Opus 4.5 for Max and Pro users (and we'll upgrade to top reasoning models as they become available). | Feb 12/13, 2026, heading 'Deep Research with Opus 4.6': Perplexity Deep Research now runs on Opus 4.6, further improving the state-of-the-art results on internal and external benchmarks that we announced just last week.", "artifact_url": "https://www.perplexity.ai/changelog/what-we-shipped---february-6th-2026 and https://www.perplexity.ai/changelog/what-we-shipped---february-13-2026", "our_archive": "ledger/archives/2026-02/perplexity/de15284d-www-perplexity-ai-changelog-what-we-shipped-february-6th-2026/20260725T151102Z and ledger/archives/2026-02/perplexity/d0f3c7d5-www-perplexity-ai-changelog-what-we-shipped-february-13-2026/20260725T151622Z", "bundle_sha256": "b3daf9f7c3c1ade34fe1455448f83748d41a2502db4bf2548539611d86ba8505 (Feb 6 page) and 3618ad72582f88f8c5b26e5150bb83a2e0a90e496ec82c61a1632d7f6be0b458 (Feb 13 page)", "source": "reported", "date_note": "2026-02-05 (before-state) / 2026-02-12 (after-state)" }, { "id": "E-2026-022", "date": "2026-02-05", "vendor": "Perplexity", "headline": "The enhanced memory engine's recall rate rose from 77% to 95%", "what_plain": "Perplexity's own changelog states its memory engine's recall rate improved from 77% to 95% in its own tests, in the same dated entry -- an explicit before/after for the same named feature.", "quote": "Memory quality over quantity: Our tests found that the enhanced memory engine was able to recall important information in 95% of cases (up from 77% previously), while makes half as many memories. In other words, it is much more effective at remembering the right things.", "artifact_url": "https://www.perplexity.ai/changelog/what-we-shipped---february-6th-2026", "our_archive": "ledger/archives/2026-02/perplexity/de15284d-www-perplexity-ai-changelog-what-we-shipped-february-6th-2026/20260725T154059Z", "bundle_sha256": "5d37e25d4ac2a4bf1d9b62242c1cb9b8abc97d48238860f8e029090bb57e3b90", "source": "reported" }, { "id": "E-2026-015", "date": "2026-02-03", "vendor": "Anthropic", "headline": "Claude Code's /model command began running immediately instead of queuing", "what_plain": "Claude Code's own release notes state the /model command changed from being queued to executing immediately.", "quote": "Changed /model to execute immediately instead of being queued", "artifact_url": "https://github.com/anthropics/claude-code/releases/tag/v2.1.30", "our_archive": "ledger/archives/2026-02/anthropic/3f5648c2-github-com-anthropics-claude-code-releases-tag-v2-1-30/20260725T151525Z", "bundle_sha256": "143031aef9ffbf7b22f70aaf362decf301e8de6fefbf52164b9ee261c73f3eff", "source": "reported" }, { "id": "E-2026-019", "date": "2026-02-03", "vendor": "OpenAI", "headline": "GPT-5.2 Thinking's Standard mode thinking time was shortened again", "what_plain": "OpenAI's own help-center thinking-time update log states a further reduction to GPT-5.2 Thinking's Standard thinking time on Feb 3 -- an additional change on top of the Jan 10 reduction already covered by January's E-2026-003, on the same named setting.", "quote": "February 3, 2026: We made another small reduction to Standard thinking time based on testing.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-02/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T154013Z", "bundle_sha256": "c79b05947ee42e01a65055c88fa602665cc2b176c082035393d8fc1364ccb39b", "source": "reported" }, { "id": "E-2026-018", "date": "2026-02-02", "vendor": "OpenAI", "headline": "Codex CLI restored its default personality to Pragmatic", "what_plain": "OpenAI's own Codex CLI release notes show the default assistant personality was changed to \"friendly\" on Feb 2, then explicitly restored to \"Pragmatic\" three days later on Feb 5 -- the same config identifier (default personality) served two different states in one week, both stated in OpenAI's own dated releases.", "quote": "Feb 2, 2026 (rust-v0.94.0): Personality configuration is now stable: default is friendly, the config key is personality, and existing settings migrate forward. | Feb 5, 2026 (rust-v0.98.0): Restored the default assistant personality to Pragmatic across config and related tests/UI snapshots.", "artifact_url": "https://github.com/openai/codex/releases/tag/rust-v0.94.0 and https://github.com/openai/codex/releases/tag/rust-v0.98.0", "our_archive": "ledger/archives/2026-02/openai/57734870-github-com-openai-codex-releases-tag-rust-v0-94-0/20260725T151709Z and ledger/archives/2026-02/openai/0139aeae-github-com-openai-codex-releases-tag-rust-v0-98-0/20260725T151747Z", "bundle_sha256": "44ffef745c618ec245796cbb7406977c88ecc20250af5e764b9fc5eec6f201f5 (Feb 2 release) and 60244ddc7e0c560ac82d223f98f5d33bc67da38fafd570b7beecbd2fc9de1f55 (Feb 5 release)", "source": "reported", "date_note": "2026-02-02 (changed to friendly) / 2026-02-05 (restored to Pragmatic)" }, { "id": "E-2026-023", "date": "2026-02-01", "vendor": "xAI", "headline": "Grok-4-fast models' context limitation above 120k tokens was mitigated", "what_plain": "xAI's own status feed states that requests to grok-4-fast-reasoning and grok-4-fast-non-reasoning over 120k tokens of context had problems, and two days later states the limitation was mitigated -- the same named models served degraded behaviour on long-context requests, then a fixed state, both stated in xAI's own dated incident record.", "quote": "Title: [API (us-east-1.api.x.ai)] grok-4-fast-reasoning and grok-4-fast-non-reasoning context size limitation. Mon, 02 Feb 2026 17:37:35 GMT -- Investigating: We are currently seeing issues on requests where context length is over 120k tokens. Currently investigating root cause. Tue, 03 Feb 2026 14:02:23 GMT -- Resolved: The context length limitation has been mitigated.", "artifact_url": "https://status.x.ai/api-us-east-1/INC686ffa19 (per status.x.ai/feed.xml)", "our_archive": "ledger/archives/2026-02/xai/431a5a63-status-x-ai-feed-xml/20260725T150956Z", "bundle_sha256": "4c382f1618f9f77bd7ab2db3b84ffd8e01df4f853997778dc20ed1270042d888", "source": "reported", "date_note": "2026-02-01 to 2026-02-03" }, { "id": "E-2026-001", "date": "2026-01-29", "vendor": "Anthropic", "headline": "Structured outputs became generally available for Claude Sonnet, Opus and Haiku 4.5", "what_plain": "Anthropic announced structured outputs went GA on Sonnet/Opus/Haiku 4.5: a parameter moved, the beta header requirement was dropped, latency described as improved.", "quote": "Structured outputs are now generally available on the Claude API for Claude Sonnet 4.5, Claude Opus 4.5, and Claude Haiku 4.5. GA includes expanded schema support, improved grammar compilation latency, and a simplified integration path with no beta header required. The `output_format` parameter has moved to `output_config.format`. Existing beta users can continue using the beta header during the transition period. Structured outputs remain in public beta on Amazon Bedrock and Microsoft Foundry.", "artifact_url": "https://platform.claude.com/docs/en/release-notes/overview", "our_archive": "ledger/archives/2026-01/anthropic/4e661726-platform-claude-com-docs-en-release-notes-overview/20260725T135428Z", "bundle_sha256": "7febb4977f35b20668c002654724909c69ea77cd82559a0eff7c201136c92da6", "source": "reported" }, { "id": "E-2026-007", "date": "2026-01-27", "vendor": "Google", "headline": "Gemini 3 became the default model behind AI Overviews", "what_plain": "Google's Search blog stated Gemini 3 became the default model behind AI Overviews, and that users can now ask follow-up questions directly from an AI Overview.", "quote": "Gemini 3 is now the default model for AI Overviews, giving you better AI responses. You can now ask follow-up questions from AI Overviews and start a conversation. This new experience provides quick answers or deeper conversations, as you need.", "artifact_url": "https://blog.google/products-and-platforms/products/search/ai-mode-ai-overviews-updates/", "our_archive": "ledger/archives/2026-01/google/e2b28a41-blog-google-products-and-platforms-products-search-ai-mode-ai-overviews-updates/20260725T105051Z", "bundle_sha256": "e382d294aa2ec2ac248b9f9d79406c0f066299298d55960bbbccf7f70c085ae0", "source": "reported", "date_note": "2026-01-27 (exact byline date not relocated this pass -- see note)" }, { "id": "E-2026-002", "date": "2026-01-22", "vendor": "OpenAI", "headline": "GPT-5.2 Instant's default personality began adapting its tone more conversationally", "what_plain": "OpenAI updated GPT-5.2 Instant's default personality via a system-prompt change: more conversational, adapts tone contextually.", "quote": "5.2 Personality System Prompt Update. We're updating GPT-5.2 Instant's default personality to be more conversational and better at adapting its tone contextually, making exchanges feel smoother and more natural. You can still select a different base style and tone for ChatGPT, along with tuning characteristics like warmth and emoji use, within the Personalization menu in settings.", "artifact_url": "https://help.openai.com/en/articles/6825453-chatgpt-release-notes", "our_archive": "ledger/archives/2026-01/openai/3930e8de-help-openai-com-en-articles-6825453-chatgpt-release-notes/20260725T140523Z", "bundle_sha256": "1e71a0703d389f325960a7c142dc22852f08d3dbeba8793e44dd984c10a64b4a", "source": "reported" }, { "id": "E-2026-006", "date": "2026-01-21", "vendor": "Google", "headline": "The gemini-pro-latest and gemini-flash-latest aliases now point to Gemini 3", "what_plain": "Google's Gemini API changelog logged that the gemini-pro-latest and gemini-flash-latest aliases were repointed from their prior targets to the Gemini 3 preview models, with the alias names unchanged.", "quote": "January 21, 2026. Changed the latest aliases: gemini-pro-latest switched to gemini-3-pro-preview. gemini-flash-latest switched to gemini-3-flash-preview.", "artifact_url": "https://ai.google.dev/gemini-api/docs/changelog", "our_archive": "ledger/archives/2026-01/google/a24fd97e-ai-google-dev-gemini-api-docs-changelog/20260725T141038Z", "bundle_sha256": "b7f0edc63ae21d585463f35423c0e8bdca8a6c771f9be954e8988c4e59105c2d", "source": "reported" }, { "id": "E-2026-004", "date": "2026-01-13", "vendor": "OpenAI", "headline": "The sora-2 alias now points to a newer dated snapshot", "what_plain": "OpenAI's API changelog logged three same-day slug repoints: gpt-realtime-mini/gpt-audio-mini, sora-2, and gpt-4o-mini-tts/gpt-4o-mini-transcribe all moved to newer dated snapshots behind the same public alias.", "quote": "Jan 13 Update gpt-realtime-mini gpt-audio-mini. Updated the gpt-realtime-mini and gpt-audio-mini slugs to point to the 2025-12-15 snapshots. If you need the previous model snapshots, use gpt-realtime-mini-2025-10-06 and gpt-audio-mini-2025-10-06. Jan 13 Update sora-2. Updated the sora-2 slug to point to sora-2-2025-12-08. If you need the previous model snapshot, use sora-2-2025-10-06. Jan 13 Update gpt-4o-mini-tts gpt-4o-mini-transcribe. Updated the gpt-4o-mini-tts and gpt-4o-mini-transcribe slugs to point to the 2025-12-15 snapshots. If you need the previous model snapshots, use gpt-4o-mini-tts-2025-03-20 and gpt-4o-mini-transcribe-2025-03-20.", "artifact_url": "https://developers.openai.com/api/docs/changelog", "our_archive": "ledger/archives/2026-01/openai/49e8d057-developers-openai-com-api-docs-changelog/20260725T140916Z", "bundle_sha256": "eab2309ecef2c0aa8963036d3475758bf624fe4ba371183cc3c44f1ab9721599", "source": "reported" }, { "id": "E-2026-003", "date": "2026-01-10", "vendor": "OpenAI", "headline": "GPT-5.2 Thinking's accidentally lowered Extended thinking time was fixed", "what_plain": "OpenAI disclosed, after the fact, that it had lowered GPT-5.2 Thinking's Standard/Light thinking time on Jan 10, and that the same change unintentionally also lowered Extended thinking, which it says it later fixed.", "quote": "Update on thinking time settings for GPT-5.2 Thinking in ChatGPT (February 4, 2026). Jan 10, 2026: We lowered the Standard and Light thinking time as we observed users prefer faster responses. As part of this update, the Extended thinking setting for GPT-5.2 was unintentionally changed to be lower which we have now fixed.", "artifact_url": "https://help.openai.com/en/articles/9624314-model-release-notes", "our_archive": "ledger/archives/2026-01/openai/badf939b-help-openai-com-en-articles-9624314-model-release-notes/20260725T140728Z", "bundle_sha256": "a42c2dbced0d775d77037c9c80eeaef712011a54a70e1d833c518bbaa2a81e1b", "source": "reported", "date_note": "2026-01-10 (disclosed 2026-02-04)" }, { "id": "E-2026-005", "date": "2026-01-09", "vendor": "OpenAI", "headline": "gpt-image-1.5 stopped forcing high fidelity on edits set to low fidelity", "what_plain": "OpenAI's API changelog logged a fix for gpt-image-1.5 and chatgpt-image-latest incorrectly applying high fidelity to image edits even when the caller explicitly set low fidelity.", "quote": "Jan 9 Fix gpt-image-1.5 chatgpt-image-latest. Fixed an issue where gpt-image-1.5 and chatgpt-image-latest were incorrectly using high fidelity for image edits through /v1/images/edits, even when fidelity was explicitly set to low (the default).", "artifact_url": "https://developers.openai.com/api/docs/changelog", "our_archive": "ledger/archives/2026-01/openai/49e8d057-developers-openai-com-api-docs-changelog/20260725T140916Z", "bundle_sha256": "eab2309ecef2c0aa8963036d3475758bf624fe4ba371183cc3c44f1ab9721599", "source": "reported" }, { "id": "E-2026-010", "date": "2026-01-09", "vendor": "OpenRouter", "headline": "The Auto Router's gateway p99 latency improved by roughly 70%", "what_plain": "OpenRouter's January release spotlight stated the Auto Router now covers 58 models (including Opus 4.5) with customizable allow-lists, and that gateway p99 latency improved roughly 70%.", "quote": "Auto Router now supports 58 models including Opus 4.5, works with tool calling, and lets you customize allowed models using wildcard syntax (e.g. anthropic/*). No markup over the routed model's market price. Per-request API support included. ... 70% faster gateway. Major p99 latency improvements across the gateway. Fastest gateways in our benchmarks.", "artifact_url": "https://openrouter.ai/blog/announcements/january-release-spotlight/", "our_archive": "ledger/archives/2026-01/openrouter/e87ff0a9-openrouter-ai-blog-announcements-january-release-spotlight/20260725T105331Z", "bundle_sha256": "e23126405b5059fd4d2ed51481ee7594feeed4449595ec9582f927b47b6e6310", "source": "reported" }, { "id": "E-2026-009", "date": "2026-01-08", "vendor": "Cursor", "headline": "Hooks in the CLI began executing about 10 times faster", "what_plain": "Cursor's CLI changelog logged Ctrl+D and Shift+Enter keybinding semantics changing, plus hooks running roughly 10x faster.", "quote": "Ctrl+D now follows standard shell behavior, requiring a double-press to exit. Shift+Enter now inserts a newline instead of submitting, making it easier to write multi-line prompts. Hooks now execute in parallel with merged responses, improving performance for projects with multiple hook scripts. Hooks execution latency reduced by 10x.", "artifact_url": "https://cursor.com/changelog/cli-jan-08-2026", "our_archive": "ledger/archives/2026-01/cursor/39dd3c83-cursor-com-changelog-cli-jan-08-2026/20260725T141157Z", "bundle_sha256": "b33ed848e4ad6c3c48caa85f377060e26145df6d2536948561e2580345bf2e9e", "source": "reported" } ] }