FrontierMath-2025-02-28-Private
Integrity rank #6 of 65 · 35 models scored · top score 68.42 · Muse Spark
unknown
Strongest on discrimination, weakest on freshness.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Muse Spark | 68.42 | reproduced· optimizedT1 | 2026-04-08 |
| 2 | Gemini 3 Pro | 65.96 | reproduced· optimizedT1 | 2025-11-18 |
| 3 | Claude Sonnet 4.6 | 56.84 | reproduced· optimizedT1 | 2026-02-17 |
| 4 | GPT-5.1 | 54.45 | reproduced· optimizedT1 | 2025-11-13 |
| 5 | Kimi K2.5 | 48.95 | reproduced· optimizedT1 | 2026-01-27 |
| 6 | Qwen 3.6 Max (Preview) | 40.53 | reproduced· optimizedT1 | 2026-04-20 |
| 7 | DeepSeek-V3.2 | 38.77 | reproduced· optimizedT1 | 2025-12-01 |
| 8 | Kimi K2 Thinking | 37.55 | reproduced· optimizedT1 | 2025-11-06 |
| 9 | Qwen 3.5 Plus (hosted 397B-A17B) | 36.9 | reproduced· optimizedT1 | 2026-02-16 |
| 10 | Grok 4 | 34.48 | reproduced· optimizedT1 | 2025-07-09 |
| 11 | GLM-5 | 28.83 | reproduced· optimizedT1 | 2026-02-11 |
| 12 | Qwen3-235B-A22B-Thinking (Jul 2025) | 14.88 | reproduced· optimizedT1 | 2025-07-25 |
| 13 | Claude Haiku 4.5 | 10.36 | reproduced· optimizedT1 | 2025-10-15 |
| 14 | Grok-3 mini | 10.28 | reproduced· optimizedT1 | 2025-04-09 |
| 15 | Gemini 2.5 Flash (Jun 2025) | 8.5 | reproduced· optimizedT1 | 2025-06-17 |
| 16 | Claude Opus 4 | 7.86 | reproduced· optimizedT1 | 2025-05-22 |
| 17 | Claude 3.7 Sonnet | 7.26 | reproduced· optimizedT1 | 2025-02-24 |
| 18 | Claude Sonnet 4 | 7.26 | reproduced· optimizedT1 | 2025-05-22 |
| 19 | Grok 3 | 6.65 | reproduced· optimizedT1 | 2025-04-09 |
| 20 | GLM-4.7 | 4.28 | reproduced· optimizedT1 | 2025-12-22 |
| 21 | Claude 3.5 Sonnet (October 2024) | 3.63 | reproduced· optimizedT1 | 2024-10-22 |
| 22 | DeepSeek-V3 | 3.02 | reproduced· optimizedT1 | 2024-12-26 |
| 23 | Gemini 2.0 Flash (Feb 2025) | 3.02 | reproduced· optimizedT1 | 2025-02-05 |
| 24 | o1-mini | 3.02 | reproduced· optimizedT1 | 2024-09-12 |
| 25 | Claude 3.5 Sonnet | 1.81 | reproduced· optimizedT1 | 2024-06-20 |
| 26 | Qwen2.5-Max | 1.81 | reproduced· optimizedT1 | 2025-01-25 |
| 27 | GPT-4.1 nano | 1.81 | reproduced· optimizedT1 | 2025-04-14 |
| 28 | Grok-2 (Dec 2024) | 1.21 | reproduced· optimizedT1 | 2024-12-12 |
| 29 | Llama 4 Maverick | 1.21 | reproduced· optimizedT1 | 2025-04-05 |
| 30 | Mistral Medium 3 | 0.61 | reproduced· optimizedT1 | 2025-05-07 |
| 31 | Claude 3.5 Haiku | 0.6 | reproduced· optimizedT1 | 2024-10-22 |
| 32 | Mistral Large 2 (Nov 2024) | 0.6 | reproduced· optimizedT1 | 2024-11-18 |
| 33 | GPT-4o (Nov 2024) | 0.6 | reproduced· optimizedT1 | 2024-11-20 |
| 34 | Gemini 1.5 Flash (Sep 2024) | 0 | reproduced· optimizedT1 | 2024-09-24 |
| 35 | Llama 4 Scout | 0 | reproduced· optimizedT1 | 2025-04-05 |
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 35 models, spanning a 68.42-point gap between the best and worst performers. That wide gap supports treating the rank ordering as real separation, since models are not bunched closely enough for small score differences to be dismissed as noise.
3 cited facts
Its Benchmark Integrity Index score is 98, with grade A and rank 6 out of 65 tracked benchmarks.
4 cited facts
This benchmark is not saturated, with a top score of 68.42 and 31.58 points of headroom to the ceiling. Only one model clusters near the top, within a couple of points. Not being saturated means there is still genuine room to separate the field at the top, so small differences among top models can signal real capability gaps rather than noise.
5 cited facts
For this benchmark, scores are directly comparable under a consistent harness, and the test set is held out rather than public. In practice, consistent comparability supports head-to-head comparisons among models evaluated under the same disclosed harness, while results from different harnesses would not be directly comparable, and the held-out test set means high scores warrant less contamination scrutiny than a public set would, though the contamination history is unknown. The harness report includes model, model version, performance, optimized flag, score source, release date, and model date, so scores should be interpreted as task performance under that disclosed harness rather than deployed capability.
7 cited facts
It is not documented what this benchmark measures, so the specific capability or behavior it targets cannot be stated from the available record. The construction of the test is likewise undocumented: there is no recorded account of how items were authored, sourced, or assembled. The benchmark's key assumptions are not documented, and no human baseline is recorded, leaving the conditions under which its results are meant to be interpreted unstated. The sharpest caveat is that because the limitations are undocumented, a reader has no stated guidance on what this benchmark cannot tell them or where it is most likely to mislead.
5 cited facts
FrontierMath-2025-02-28-Private: FrontierMath-2025-02-28-Private as reported in Epoch AI's Capabilities Index CSV.
Muse Spark leads FrontierMath-2025-02-28-Private at 68.42 — independently reproduced. The full leaderboard above lists every recorded measurement, not just the headline number.
35 models have recorded scores on FrontierMath-2025-02-28-Private, spanning a score spread of 68.42.
tensor.news grades FrontierMath-2025-02-28-Private A for integrity (score 98/100), ranking #6 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — FrontierMath-2025-02-28-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-2025-02-28-Private are reasonably apples-to-apples.
Every FrontierMath-2025-02-28-Private measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.