OpenAI Decisions API shows 68% accuracy on high-confidence claims
October 1, 2026
Testing of OpenAI's Decisions API, powered by GPT-6 Luna, reveals that when the model claims 99% confidence on reasoning problems, it is correct only 68% of the time. The low-latency model targets 150ms decision times for tasks like request routing and agentic branching.
HOW THIS AFFECTS YOU
●
builderYou must implement external verification for agentic actions, as high confidence scores are currently unreliable.
●
policyThe discrepancy between stated confidence and actual accuracy poses significant safety risks for autonomous systems.