ValsAI benchmark tests cost of reasoning levels in GPT-6 Astra and Opus 5.5

Vals AI @ValsAIOriginally posted 2026.09.29

Open original on X

When does more reasoning stop being useful and start becoming overhead? We ran Open AI and Anthropic’s best models: GPT-6 Astra and Claude Opus 5.5 at every reasoning level on our new benchmark, MysteryMechanism. Thinking longer delivers accuracy gains with the same model at as much as 20x the cost. Inspecting traces, the models use their effort level in very different ways.