SI USA Superintelligence in the USA Subscribe
Menu

Timeline Evaluation SIU-LDG-2026-014 SI Index +2.50

Corrected FrontierMath shows AI solving most research-level problems

Epoch AI's audited FrontierMath v2 corrected errors in 42% of problems. On its own runs, the best models solve most of Tiers 1–3 and most of Tier 4, the hardest problems.

Bar chart of FrontierMath Tier 4 version 2 scores, with Claude Fable 5 highest at 90 percent and six other models, most above half
FrontierMath Tier 4 (v2), Epoch-run, models released by June 12, 2026. Chart: SI USA, from Epoch AI data / Data CC BY 4.0 · License

FrontierMath problems are written by working mathematicians and checked automatically. The benchmark crossed the 25% level on Epoch’s own runs in mid-2025. By October 2026 OpenAI’s newest models scored above 90% on both sets, which means the benchmark is close to saturated as a measure of this milestone.

Index treatment. Research-level mathematics moves to 0.5. The next level needs a new theorem accepted by working mathematicians.