New
AI & RoboticsMore in AI & Robotics→
OpenAI’s ARC-AGI-3 claim turns on the test setup
Key Takeaways
- OpenAI reported a 38.3% ARC-AGI-3 score for GPT-5.6 Sol using its own API features.
- The model scored 7.8% in the benchmark’s official harness, where reasoning is not retained between steps.
- ARC Prize said general-purpose API settings may be fair to use if they are disclosed, but noted parity issues across providers.
DE
DT Editorial Team··via the-decoder.com