The annotation on arc-agi-3 is this: > OpenAI's own evaluation notes say Astra uses the company's Responses API harness, while comparison models can operate under different configurations.

With this configuration gpt-5.6-sol was able to reach 38,3%. So this is misleading.