All Benchmarks/Math

FrontierMath Tier 4 (2025-07)

Older versionHigher is better

The old version 1 (48 private problems) of Tier 4, the hardest research-level set in Epoch AI's FrontierMath. The AI can use Python and must give an exact answer. It was superseded by version 2 in June 2026, is no longer updated, and is not comparable with version 2 scores. Scores run from 0 to 100%. Higher is better.

A newer version is available: FrontierMath Tier 4 v2

Other versions:FrontierMath Tiers 1-3 (2025-02)FrontierMath: Erdős ProblemsFrontierMath Tier 4 v2FrontierMath Tiers 1-3 v2
Top score31.3%Claude Opus 4.8
Models tested29
Last updated2026.06.08
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 1
    Anthropic
    Claude Opus 4.8
    2026.05.27 · max
    31.3%
  2. 2
    OpenAI
    GPT-5.4
    2026.03.05 · xhigh
    27.1%
  3. 3
    Anthropic
    Claude Opus 4.7
    2026.04.16 · xhigh
    22.9%
  4. 4
    Anthropic
    Claude Opus 4.6
    2026.02.04 · max
    22.9%
  5. 5
    Google
    Gemini 3.1 Pro
    2026.02.19
    16.7%
  6. 6
    Meta
    Muse Spark
    2026.04.08
    14.6%
  7. 7
    Google
    Gemini 3.5 Flash
    2026.05.19 · high
    14.6%
  8. 8
    Moonshot AI
    Kimi K2.6
    2026.04.20
    14.6%
  9. 9
    OpenAI
    GPT-5
    2025.08.07 · high
    12.5%
  10. 9
    Z.ai
    GLM-5.1
    2026.04.07
    12.5%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.