GPT-6 Astra hits 99.9% on ARC-AGI-3, or 62.7% on a neutral harness
ARC Prize scored Astra both ways: 99.9% with OpenAI's adapter, 62.7% on its own standard harness. Same model, a 37-point gap.
OpenAI launched GPT-6 Astra leading with 99.9% on ARC-AGI-3. ARC Prize, the group that built the benchmark, ran the same model on its own provider-neutral Standard harness and got 62.7% at max reasoning effort.
The difference is the setup around the model. OpenAI's Provider Adapter harness carries Astra's hidden reasoning state from one request to the next and compacts long conversations. The Standard harness only lets the model keep notes it chooses to write. Under the adapter, every reasoning level scores between 96.7% and 99.9%. On the Standard harness, scores run from 17.5% at low effort to 62.7% at max. The neutral run was also more expensive: $26,098 against $18,817 for the 99.9% run.
ARC Prize isn't calling either number wrong. It will list both results on its leaderboard, each labelled with its harness, and it states plainly that it is not claiming Astra is AGI. Artificial Analysis puts Astra at 61 on its Intelligence Index, the same as GPT-5.6 Sol, while Astra's API input price is double Sol's.
Why it matters: a benchmark score now partly measures the vendor's harness, so check which harness produced a number before comparing models.
Sources
Independent coverage
- OpenAI's GPT-6 Astra on ARC-AGI-3 arcprize.org
- Astra's AGI benchmark reads 99.9% or 62.7% implicator.ai
Written by an AI pipeline from the sources above. Methodology · Report an error
The stories that matter for people who build with AI, with the source for every claim. Free, no spam.
Feed, daily deep-dive and bytes — readable offline, with push alerts for the topics you follow.