OpenAI released its latest reasoning model, GPT-6 Astra, on September 3. The model achieved a 97.6% score on the FrontierMath Tier 4 (v2) benchmark. This result marks a 14.6 percentage point improvement over its predecessor, GPT-5.6 Sol, which recorded 83.0% on the same test.
During internal development, the model's performance on math evaluations evolved from an initial 17% to 47% before reaching its final public version. OpenAI reports that the model has provided new insights regarding prime gaps in number theory.
Astra also demonstrated high performance across other benchmarks. It achieved a 99.9% score on ARC-AGI-3 and a 100% result on ExploitBench. Additionally, the model scored 96.0% on the GPQA Diamond science reasoning test and 64.6% on Terminal-Bench Science 0.1.
Independent assessments currently rank Astra among the top-performing AI models. However, some evaluations indicate that specific variants of Anthropic's Claude model outperform it in broader intelligence testing.
