FrontierMath
Integrity rank #10 of 61 · 48 models scored · top score 89.12 · GPT-5.6 Sol
Research-level novel math, far beyond competition
Strongest on discrimination, weakest on saturation headroom.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | GPT-5.6 Sol | 89.12 | reproducedT1 | 2026-07-09 |
| 2 | GPT-5.5 Pro | 87.72 | reproducedT1 | 2026-04-23 |
| 3 | Claude Fable 5 | 87.02 | reproducedT1 | 2026-06-09 |
| 4 | GPT-5.6 Terra | 85.96 | reproducedT1 | 2026-07-09 |
| 5 | Claude Opus 5 | 85.61 | reproducedT1 | 2026-07-24 |
| 6 | GPT-5.5 | 85.26 | reproducedT1 | 2026-04-23 |
| 7 | GPT-5.4 Pro | 82.46 | reproducedT1 | 2026-03-05 |
| 8 | GPT-5.6 Luna | 82.11 | reproducedT1 | 2026-07-09 |
| 9 | Claude Opus 4.8 | 80 | reproducedT1 | 2026-05-28 |
| 10 | GPT-5.4 | 78.6 | reproducedT1 | 2026-03-05 |
| 11 | Qwen 3.8 Max | 74.74 | reproducedT1 | 2026-07-19 |
| 12 | GPT-5.2 Pro | 74 | reproducedT1 | 2025-12-11 |
| 13 | Kimi K3 | 72.18 | reproducedT1 | 2026-07-16 |
| 14 | Gemini 3.7 Flash | 71.58 | reproducedT1 | 2026-08-13 |
| 15 | Claude Opus 4.7 | 70.18 | reproducedT1 | 2026-04-16 |
| 16 | GPT-5.2 | 67.4 | reproducedT1 | 2025-12-11 |
| 17 | Grok 4.6 | 65.96 | reproducedT1 | 2026-08-12 |
| 18 | Claude Opus 4.6 | 65.96 | reproducedT1 | 2026-02-05 |
| 19 | Claude Sonnet 5 | 65.61 | reproducedT1 | 2026-06-30 |
| 20 | Qwen3.7-Max | 64.56 | reproducedT1 | 2026-05-19 |
| 21 | Gemini 3.5 Flash | 62.81 | reproducedT1 | 2026-05-19 |
| 22 | Gemini 3.1 Pro | 59.65 | reproducedT1 | 2026-02-19 |
| 23 | GLM-5.2 | 59.21 | reproducedT1 | 2026-06-16 |
| 24 | Gemini 3.6 Flash | 58.95 | reproducedT1 | 2026-07-21 |
| 25 | DeepSeek V4 Flash 0731 | 57.54 | reproducedT1 | 2026-07-31 |
| 26 | Kimi K2.6 | 57.19 | reproducedT1 | 2026-04-20 |
| 27 | Grok 4.5 | 57.19 | reproducedT1 | 2026-07-08 |
| 28 | GPT-5 Pro | 55.79 | reproducedT1 | 2025-10-07 |
| 29 | GPT-5 | 55.44 | reproducedT1 | 2025-08-07 |
| 30 | Kimi K2.7 Code | 54.04 | reproducedT1 | 2026-06-12 |
| 31 | Gemini 3 Flash | 51.23 | reproducedT1 | 2025-12-17 |
| 32 | GPT-5.4 Mini | 51.23 | reproducedT1 | 2026-03-17 |
| 33 | GPT-5 mini | 46.67 | reproducedT1 | 2025-08-07 |
| 34 | Inkling-Small | 46.32 | reproducedT1 | 2026-07-30 |
| 35 | DeepSeek-V4-Pro | 45.26 | reproducedT1 | 2026-04-24 |
| 36 | Grok 4.20 | 44.91 | reproducedT1 | 2026-02-17 |
| 37 | GPT-5.4 Nano | 44.91 | reproducedT1 | 2026-03-17 |
| 38 | Grok 4.3 Beta | 42.81 | reproducedT1 | 2026-04-17 |
| 39 | o4-mini | 36.14 | reproducedT1 | 2025-04-16 |
| 40 | Claude Opus 4.5 | 34.39 | reproducedT1 | 2025-11-24 |
| 41 | Inkling | 33.33 | reproducedT1 | 2026-07-15 |
| 42 | GPT-5.5 Instant | 26.32 | reproducedT1 | 2026-05-05 |
| 43 | Gemini 3.5 Flash-Lite | 25.96 | reproducedT1 | 2026-07-21 |
| 44 | Gemini 2.5 Pro (Jun 2025) | 24.56 | reproducedT1 | 2025-06-05 |
| 45 | Claude Sonnet 4.5 | 23.86 | reproducedT1 | 2025-09-29 |
| 46 | GPT-5 nano | 20 | reproducedT1 | 2025-08-07 |
| 47 | o3-mini | 18.6 | reproducedT1 | 2025-01-31 |
| 48 | Claude Opus 4.1 | 12.63 | reproducedT1 | 2025-08-05 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark scores 48 models, with a 76.49-point gap between the best and worst scores. That wide spread supports real separation between models, so rank differences are meaningful rather than likely noise.
2 cited facts
The benchmark's integrity score is 99, its integrity grade is A, and it ranks 10th out of 61 benchmarks in benchmark universe size. The weakest component of the integrity breakdown is saturation, which costs the reader a crowded ceiling among top scores. This integrity score should be read as the benchmark's health as a discriminator of frontier models under a disclosed harness, not as any model's capability.
6 cited facts
This benchmark is not saturated. The top score is 89.12, leaving 10.88 points of headroom, and only two models cluster within a couple of points of the top. Because the field is not saturated, there is still genuine room to separate the top models, so small differences in scores remain meaningful.
6 cited facts
Because this benchmark uses a consistent harness, scores are directly comparable across runs, so task performance can be compared head-to-head without adjustment. The test set is held out, reducing the risk that high scores reflect contamination, though the funder's previous access to some tier-4 items under a verbal no-train agreement and a subsequent funding scandal introduce qualitative uncertainty. Performance is sensitive to harness choices such as single-shot prompting, Python execution, and heavy test-time compute, so scores represent task performance under this disclosed harness, not deployed capability.
4 cited facts
This benchmark measures research-level novel mathematics far beyond competition, so its scores are intended to reflect original discovery rather than routine problem-solving. It was built by more than sixty mathematicians, with a held-out set of auto-checkable problems arranged in difficulty tiers; the design assumes that novelty plus automatic verification rules out contamination and makes correct answers evidence of discovery. The set excludes proofs, is very small in size, and a correct answer can hide an invalid method, while access to the problems has also been controversial. The sharpest caveat is that an auto-verified correct answer does not certify the reasoning behind it, so a strong score should not be read as proof-level mathematical competence.
6 cited facts
FrontierMath-Tiers-1-3-v2-Private (FrontierMath): FrontierMath (private v2, tiers 1-3) — original, unpublished research-level mathematics problems commissioned by Epoch AI to resist memorization and contamination; scored as accuracy (%). The private held-out set keeps the test set uncontaminated.
GPT-5.6 Sol leads FrontierMath-Tiers-1-3-v2-Private at 89.12 — independently reproduced. The full leaderboard above lists every recorded measurement, not just the headline number.
48 models have recorded scores on FrontierMath-Tiers-1-3-v2-Private, spanning a score spread of 76.49.
tensor.news grades FrontierMath-Tiers-1-3-v2-Private A for integrity (score 99/100), ranking #10 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — FrontierMath-Tiers-1-3-v2-Private still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is held-out, which lowers — but does not eliminate — the risk that training data leaked into the questions.
Scores are reported under a consistent harness, so comparisons on FrontierMath-Tiers-1-3-v2-Private are reasonably apples-to-apples.
Every FrontierMath-Tiers-1-3-v2-Private measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/frontiermath.