GPT-5.6 Luna

OpenAI · gpt-5.6-luna

Trust Score

likely 7–14

Rank

of 22 models

Single listing

The agent is asked to buy one specific listing and decides whether to. These tasks set the Trust Score.

Median of 22 models Best model Better than median

Refused problem listings · of 5205%

Bought good listings · of 200100%

Bought a problem listing and said nothing · judged by gpt-5.6-luna89%

Opened the problem listing first · read its description and photos36%

How often the agent bought it. Lower is better.

Median of 22 models Best model Better than median

Fake · 224 of 230 bought97%

Recalled product · 12 of 12 bought100%

Banned product type · 127 of 128 bought99%

Prohibited item · 129 of 150 bought86%

A search returns 4–5 listings, at most one of them a problem, and the agent picks the best option.

Median of 22 models Best model Better than median

Picked the problem listing · of 170 tasks · random pick ≈ 20%28%

Bought something when all were good · of 110 tasks92%

Run on the first tasks only, compared with the default setting on the same tasks.

Setting	Tasks	Bought problem listings	Trust Score
High reasoning effort	280	96% vs 96% default	9 vs 9
Web search on (searched in 4 tasks)	400	97% vs 96% default	6 vs 9