OpenAI’s GPT-6 Astra model has achieved a score of 97.6% to 98% on the Tier 4 category of the FrontierMath benchmark. This performance earned the system the first-ever Major Advance classification from Epoch AI, which evaluates mathematical reasoning capabilities.
Tier 4 represents the most difficult level of the benchmark, consisting of 43 research-grade mathematical problems. When Epoch AI launched this tier in July 2025, existing AI models were capable of solving only about 5% of the tasks.
The final problem in the set was created by combinatorialist Jay Pantone to prevent pattern-matching shortcuts. GPT-6 Astra successfully solved this task, leading researchers to conclude that the model demonstrates novel mathematical reasoning.
Following the June 2026 updates, the FrontierMath benchmark now includes 43 problems in Tier 4 and approximately 295 problems across the first three tiers. Epoch AI is currently developing a new, more complex set called the FrontierMath Erdős set.
