GPT-6.1 Sol Maxes Out Research Level FrontierMath Tier 4

The latest mathematical AI has achieved a perfect score on a testing suite previously deemed too difficult for most human researchers. The model, known as GPT-6.1 Sol, reached this performance ceiling after a development period of 10 months.

The agent saturated the FrontierMath Tier 4 benchmark, which was considered exceptionally difficult even for research-level mathematics last year. In a separate development, an AI named Astra solved 5 Erdős problems, completing 2 of them under the assigned budget.