FrontierMath-Tier-4-2025-07-01-Private
Integrity rank #18 of 65 · 26 models scored · top score 31.25 · Gemini 3 Pro
unknown
Strongest on contamination resistance, weakest on discrimination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Gemini 3 Pro | 31.25 | reproducedT1 | 2025-11-18 |
| 2 | Muse Spark | 24.33 | reproducedT1 | 2026-04-08 |
| 3 | GLM-5.1 | 20.83 | reproducedT1 | 2026-04-07 |
| 4 | GPT-5.1 | 20.83 | reproducedT1 | 2025-11-13 |
| 5 | Qwen 3.6 Plus | 13.89 | reproducedT1 | 2026-03-31 |
| 6 | Claude Sonnet 4.6 | 13.83 | reproducedT1 | 2026-02-17 |
| 7 | Kimi K2.5 | 7 | reproducedT1 | 2026-01-27 |
| 8 | Gemini 2.5 Flash (Jun 2025) | 6.94 | reproducedT1 | 2025-06-17 |
| 9 | Claude Opus 4 | 6.94 | reproducedT1 | 2025-05-22 |
| 10 | Qwen 3.6 Max (Preview) | 6.94 | reproducedT1 | 2026-04-20 |
| 11 | DeepSeek-V3.2 | 3.5 | reproducedT1 | 2025-12-01 |
| 12 | GLM-5 | 3.5 | reproducedT1 | 2026-02-11 |
| 13 | Grok 4 | 3.47 | reproducedT1 | 2025-07-09 |
| 14 | Claude Haiku 4.5 | 3.47 | reproducedT1 | 2025-10-15 |
| 15 | Qwen 3.5 Plus (hosted 397B-A17B) | 3.47 | reproducedT1 | 2026-02-16 |
| 16 | o3 | 3.47 | reproducedT1 | 2025-04-16 |
| 17 | Claude 3.5 Sonnet | 0 | reproducedT1 | 2024-06-20 |
| 18 | Claude 3.5 Sonnet (October 2024) | 0 | reproducedT1 | 2024-10-22 |
| 19 | Grok 3 | 0 | reproducedT1 | 2025-04-09 |
| 20 | Kimi K2 Thinking | 0 | reproducedT1 | 2025-11-06 |
| 21 | Qwen 3.5 Flash (hosted 35B-A3B) | 0 | reproducedT1 | 2026-02-25 |
| 22 | Qwen 3.6 Flash | 0 | reproducedT1 | 2026-04-27 |
| 23 | Qwen3-235B-A22B-Thinking (Jul 2025) | 0 | reproducedT1 | 2025-07-25 |
| 24 | Claude Sonnet 4 | 0 | reproducedT1 | 2025-05-22 |
| 25 | GLM-4.7 | 0 | reproducedT1 | 2025-12-22 |
| 26 | GPT-4.1 | 0 | reproducedT1 | 2025-04-14 |
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.
Across the models evaluated, 26 are scored on this benchmark, and the gap between the highest and lowest scoring models is 31.25 points. A spread of that size points to substantial separation in performance, so rank differences among these models are more likely to reflect genuine differences in capability rather than noise from closely clustered scores.
3 cited facts
This benchmark's Benchmark Integrity Index is 93, with grade A and a rank of 16th out of 62 benchmarks. The weakest component in the integrity breakdown is discrimination, which limits how finely the benchmark can separate frontier models under its disclosed harness, so small score gaps should be interpreted cautiously. Read the score 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 best recorded score is 31.25 points, leaving about 68.75 points of headroom before the ceiling is reached. Only a single model currently clusters within a couple of points of that top score. Because the benchmark is not saturated, there is still genuine room to separate the field at the top, so gaps between the leading models can be read as meaningful rather than as noise.
5 cited facts
For this benchmark, harness comparability is consistent, so reported scores are directly comparable across entries that use the same evaluation setup. The test set is held out, which means high scores warrant less suspicion of contamination than a public test set would, although contamination history is unknown and cannot be ruled out. Because comparability is tied to the harness, scores from different harnesses should not be compared head-to-head, and all figures should be read as task performance under a disclosed harness rather than deployed capability.
5 cited facts
What this benchmark measures is not documented in the available record, so its target capability or quantity cannot be stated here. How the test was built is likewise undocumented, meaning no description of its task-construction process or design rationale can be given. The assumptions the benchmark rests on and the limitations it acknowledges are also undocumented, leaving its interpretive boundaries unstated. The sharpest caveat a reader should hold is that no human baseline is documented, so performance on this benchmark cannot be anchored to a human reference point.
6 cited facts
FrontierMath-Tier-4-2025-07-01-Private: FrontierMath-Tier-4-2025-07-01-Private as reported in Epoch AI's Capabilities Index CSV.
Gemini 3 Pro leads FrontierMath-Tier-4-2025-07-01-Private at 31.25 — independently reproduced. The full leaderboard above lists every recorded measurement, not just the headline number.
26 models have recorded scores on FrontierMath-Tier-4-2025-07-01-Private, spanning a score spread of 31.25.
tensor.news grades FrontierMath-Tier-4-2025-07-01-Private A for integrity (score 93/100), ranking #18 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — FrontierMath-Tier-4-2025-07-01-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-Tier-4-2025-07-01-Private are reasonably apples-to-apples.
Every FrontierMath-Tier-4-2025-07-01-Private measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.