OpenAI's GPT-6 Astra clears FrontierMath Tier 4's last unsolved problem
The AMW Read
Updates the OpenAI case study with Tier 4 saturation—a designed-hard math wall collapsing in ~14 months—while Erdős scores keep the capability claim bounded.
OpenAI's GPT-6 Astra clears FrontierMath Tier 4's last unsolved problem
According to a QubitAI report dated September 12, 2026, OpenAI's GPT-6 Astra scored 97.6% on FrontierMath Tier 4 and solved the final problem no prior model had cracked. Epoch AI concluded the tier is saturated under its cumulative rule: every remaining Tier 4 item has now been solved at least once across models and attempts. FrontierMath was launched November 7, 2024 with Fields Medalists including Terence Tao among contributors; Tier 4 was added July 11, 2025 when peak scores sat near 5%. After a June 2026 v2 audit that fixed 12 problems, removed 7, and left 43, GPT-5.6 Sol reached 83.0%, Claude Fable 5 hit 90.2%, then Astra 97.6%. Jay Pantone of Marquette University said Astra's solution closely matched his own rather than a numerical shortcut.
Saturating a research-grade math wall built to resist years of progress in roughly fourteen months weakens static hard benchmarks as lasting capability moats among frontier labs. The same week OpenAI has been pushing contested high-profile math claims—including Millennium Prize and agent-swarm results covered previously by AI Market Watch—this Epoch milestone converts that campaign into a named evaluation outcome. Claude Fable 5's 90.2% shows the race remains multi-player. Per the AI Market Watch index, OpenAI matched 321 news items in the last 90 days versus 271 in the prior 90, coverage limited to pipeline-ingested sources.
Builders and investors should read Tier 4 saturation as a capability signal, not a solved-math verdict. On FrontierMath Erdős, Astra solved only 2 of 68 formalized open problems, and Epoch has already moved the bar toward genuine open problems and Lean-verified proofs. Product and eval bets that assumed research math would stay AI-hard for years need a shorter obsolescence clock.
