Exchange

Decide

Apply the rules to a refund request

Pick · 10 Nov

Highest (right answers per dollar) among models at least 90% right; no pick if none qualify.

GPT-6 Luna

6,100 right answers per $1

$0.16 per 1,000 right

Buy at OpenAI (opens OpenAI, another site)

Model comparison

How the models compare

50 items, same for every model.
ModelRight (%)Cost per 1,000 right
GPT-6 LunaPick100$0.16
GLM-5.3 Flash98$0.31
Claude Haiku 5.5100$0.33
Gemini 3.1 Flash Lite90$0.43

The test

Goes in
A refund request
Comes out
A refund decision and amount
Right means
Correct decision and amount to the cent
Checked by
A script

Models testedPreview Desk

Desk · Not open yet.

If another company’s model has higher than Desk that day, we’ll link you to the highest- model before you send a job. You will still be able to use Desk.