Why this benchmark is useful
When MATH and AIME no longer separate the frontier, this is the live math discriminator: original expert problems, Python-enabled, still far from ceiling.
FrontierMath
FrontierMath Tiers 1-3
Hard unpublished mathematics: 295 private Tiers 1-3 problems (v2) where the model writes a Python answer() after iterative reasoning and code execution.
What this does not measure
- Tier 4 research-level problems — that 43-problem expansion is a separate, harder set. Do not mix Tiers 1-3 headlines with Tier 4.
- Contest MATH / AIME skill — those suites are saturated. FrontierMath is original expert-authored problems, not recycled olympiad papers.
- Whether the derivation is human-readable — scoring is pass/fail on the submitted Python object, not a proof grade.
Coverage map
- Task family
- Math
- Format
- Private set of 295 Tiers 1-3 problems (v2, 12 Jun 2026 correction). The model may think aloud, call a stateless Python tool, and must submit an answer() function. Token cap 1,000,000.
- Scoring
- Pass rate on the private Tiers 1-3 v2 set (1 if the submitted object matches, else 0). VerdictPal rows use Epoch's best internal run per model from the public CSV.
- Maintainer
- Epoch AI
How to read the scores
Read the private Tiers 1-3 v2 pass rate, not vendor MATH-500 headlines. Epoch funds include OpenAI exclusive access to a subset — treat OpenAI rows with that conflict in mind. GPT-5.6 Sol is not in the 2026-08-15 Epoch CSV yet.
What it does not cover
- Tier 4 research-level problems — that 43-problem expansion is a separate, harder set. Do not mix Tiers 1-3 headlines with Tier 4.
- Contest MATH / AIME skill — those suites are saturated. FrontierMath is original expert-authored problems, not recycled olympiad papers.
- Whether the derivation is human-readable — scoring is pass/fail on the submitted Python object, not a proof grade.
Scores
15 rowsEvidence ledger
295 private problems after the 12 Jun 2026 correction. This is the hub primary.
1515 rows
52.4%best score
15source-checked
1sources
2026-08-15source date
Where the rows land
0255075100
Normalized to this benchmark's axis (0–100). Open the table for raw units.
Newest receipts
- GPT-5.5 Pro52.4% · Epoch AI FrontierMath Tiers 1-3 (v2) CSV
- GPT 5.551.7% · Epoch AI FrontierMath Tiers 1-3 (v2) CSV
- GPT-5.4 Pro50% · Epoch AI FrontierMath Tiers 1-3 (v2) CSV
- GPT 5.447.6% · Epoch AI FrontierMath Tiers 1-3 (v2) CSV
- Claude Opus 4.847.24% · Epoch AI FrontierMath Tiers 1-3 (v2) CSV
Scores use this benchmark's own unit and axis, not a universal quality score.
| # | Model | Release date | Score | Provenance | Trust |
|---|---|---|---|---|---|
| 1 | GPT-5.5 ProOpenAIEpoch CSV row `gpt-5.5-pro-pre-release_high`; best internal Tiers 1-3 v2 run (started 2026-04-23). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-04-23 | 52.4% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 2 | GPT 5.5OpenAIEpoch CSV row `gpt-5.5-pre-release_xhigh`; best internal Tiers 1-3 v2 run (started 2026-04-23). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-04-23 | 51.7% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 3 | GPT-5.4 ProOpenAIEpoch CSV row `gpt-5.4-pro-2026-03-05_xhigh`; best internal Tiers 1-3 v2 run (started 2026-03-06). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-03-05 | 50% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 4 | GPT 5.4OpenAIEpoch CSV row `gpt-5.4-2026-03-05_xhigh`; best internal Tiers 1-3 v2 run (started 2026-03-06). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-03-05 | 47.6% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 5 | Claude Opus 4.8AnthropicEpoch CSV row `claude-opus-4-8_max`; best internal Tiers 1-3 v2 run (started 2026-06-08). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-05-28 | 47.24% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 6 | Claude Opus 4.7AnthropicEpoch CSV row `claude-opus-4-7_xhigh`; best internal Tiers 1-3 v2 run (started 2026-04-17). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-04-16 | 43.79% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 7 | Claude Opus 4.6AnthropicEpoch CSV row `claude-opus-4-6_max`; best internal Tiers 1-3 v2 run (started 2026-02-12). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-02-05 | 40.7% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 8 | GPT-5.2OpenAIEpoch CSV row `gpt-5.2-2025-12-11_xhigh`; best internal Tiers 1-3 v2 run (started 2025-12-13). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2025-12-11 | 40.7% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 9 | Muse SparkOtherEpoch CSV row `muse-spark`; best internal Tiers 1-3 v2 run (started 2026-04-08). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-01-01 | 39% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 10 | Gemini 3.5 FlashGoogleEpoch CSV row `gemini-3.5-flash_high`; best internal Tiers 1-3 v2 run (started 2026-05-22). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-05-19 | 38.97% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 11 | Kimi K2.6MoonshotEpoch CSV row `kimi-k2.6`; best internal Tiers 1-3 v2 run (started 2026-05-07). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-04-20 | 38.97% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 12 | Gemini 3 ProGoogleEpoch CSV row `gemini-3-pro-preview`; best internal Tiers 1-3 v2 run (started 2025-11-21). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2025-11-18 | 37.6% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 13 | Gemini 3.1 ProGoogleEpoch CSV row `gemini-3.1-pro-preview`; best internal Tiers 1-3 v2 run (started 2026-02-19). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-02-19 | 36.9% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 14 | GLM-5.1ZhipuEpoch CSV row `glm-5.1`; best internal Tiers 1-3 v2 run (started 2026-05-11). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-04-07 | 33.45% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
| 15 | Claude Sonnet 4.6AnthropicEpoch CSV row `claude-sonnet-4-6_16K`; best internal Tiers 1-3 v2 run (started 2026-02-20). OpenAI funded FrontierMath. Epoch publishes a conflict-of-interest note. | 2026-02-17 | 32.4% | Epoch AI FrontierMath Tiers 1-3 (v2) CSV2026-08-15 | Source-checked |
What it covers
Scores ingested from Epoch's 2026-08-15 public CSV (internal runs). One row per atlas model, best effort variant. OpenAI funded FrontierMath; Epoch publishes a conflict-of-interest note.
Where it breaks down
Human expert time is hours to days per problem. Frontier pass rates in the Epoch CSV cluster well below 60% on Tiers 1-3 v2.
Receipts
Sources and further reading
- Epoch FrontierMath Tiers 1-3 (v2)· checked 2026-08-15
- Epoch benchmarking data download· checked 2026-08-15
- FrontierMath program· checked 2026-08-15
- Tiers 1-3 (v2)
- CSV download
Questions about this benchmark
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
What does FrontierMath measure?
Hard unpublished mathematics: 295 private Tiers 1-3 problems (v2) where the model writes a Python answer() after iterative reasoning and code execution.
What does a high FrontierMath score not prove?
A strong FrontierMath result says nothing about:
- Tier 4 research-level problems — that 43-problem expansion is a separate, harder set. Do not mix Tiers 1-3 headlines with Tier 4.
- Contest MATH / AIME skill — those suites are saturated. FrontierMath is original expert-authored problems, not recycled olympiad papers.
- Whether the derivation is human-readable — scoring is pass/fail on the submitted Python object, not a proof grade.
How is FrontierMath scored?
Pass rate on the private Tiers 1-3 v2 set (1 if the submitted object matches, else 0). VerdictPal rows use Epoch's best internal run per model from the public CSV. Task format: Private set of 295 Tiers 1-3 problems (v2, 12 Jun 2026 correction). The model may think aloud, call a stateless Python tool, and must submit an answer() function. Token cap 1,000,000.
Is FrontierMath saturated?
FrontierMath is currently marked Active in the atlas. Ceiling context: Human expert time is hours to days per problem. Frontier pass rates in the Epoch CSV cluster well below 60% on Tiers 1-3 v2.
Can FrontierMath results be contaminated by training data?
Contamination risk for FrontierMath is graded Low contamination. Treat every row on this page as a public claim with a source and a date, not as a controlled experiment.
The Pack · Editorial newsletter
New tools in your inbox. Free.
One short email when a tool ships or changes status. No tracking, no third-party analytics. Unsubscribe in one click.