OpenAI will sell you Astra, but not the system that scored 98.6% on ARC-AGI-3

OpenAI plans to sell a model called Astra, but it won't be releasing the system that achieved a remarkable 98.6% score on the ARC-AGI-3 benchmark. Investor Matt Turck, who has interviewed the creators of ARC-AGI, described Astra as "wild," with even more enthusiastic comments in parentheses. ARC Prize, the organization that developed the benchmark, ran GPT-6 Astra through its own harness, scoring 62.7%.

When the same model was run through OpenAI's Provider Adapter, with a different harness, the score increased to 98.6%. The software surrounding the model can impact both its performance and cost. ARC Prize's standard harness allows a model to carry forward its notes, while OpenAI's adapter preserves the model's reasoning state between requests, compressing longer conversations so the model can resume its thinking.

The adapter runs on documented Responses API capabilities that anyone can use, but the assembled system that scored 98.6% remains a proprietary product. OpenAI President Greg Brockman's recent claim that "we are now in the AGI era" is based on a benchmark result produced by a particular system and not on establishing that the underlying model is actually AGI.