AnalysisAI ModelsOctober 2, 2026

OpenAI Decisions API confidence scores found unreliable

Read original source →anth.us

AnthusAI tested GPT-6 Luna, the model behind OpenAI's Decisions API, on 3,600 reasoning problems: when Luna claimed 99% confidence or higher, it was right only 68% of the time. The API, announced at DevDay on September 29 in limited preview, picks one answer from a defined list and runs at 150ms per decision.

1 source

More stories today

Open the live feed