EBR-bench
Integrity rank #6 of 61 · 18 models scored · top score 50 · Claude Opus 5
unknown
Strongest on discrimination, weakest on saturation headroom. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Claude Opus 5 | 50 | reproducedT1 | 2026-07-24 |
| 2 | Claude Fable 5 | 39.52 | reproducedT1 | 2026-06-09 |
| 3 | GPT-5.6 Sol | 39.05 | reproducedT1 | 2026-07-09 |
| 4 | GPT-5.5 | 34.29 | reproducedT1 | 2026-04-23 |
| 5 | Claude Opus 4.8 | 28.57 | reproducedT1 | 2026-05-28 |
| 6 | GPT-5.4 | 25.4 | reproducedT1 | 2026-03-05 |
| 7 | GPT-5.2 | 23.02 | reproducedT1 | 2025-12-11 |
| 8 | Claude Opus 4.7 | 19.05 | reproducedT1 | 2026-04-16 |
| 9 | Gemini 3.1 Pro | 14.29 | reproducedT1 | 2026-02-19 |
| 10 | Claude Opus 4.5 | 14.29 | reproducedT1 | 2025-11-24 |
| 11 | Claude Opus 4.6 | 12.7 | reproducedT1 | 2026-02-05 |
| 12 | GPT-5 | 12.7 | reproducedT1 | 2025-08-07 |
| 13 | GLM-5.2 | 9.52 | reproducedT1 | 2026-06-16 |
| 14 | Qwen3.7-Max | 9.52 | reproducedT1 | 2026-05-19 |
| 15 | Claude Opus 4.1 | 7.94 | reproducedT1 | 2025-08-05 |
| 16 | Gemini 3.5 Flash | 4.76 | reproducedT1 | 2026-05-19 |
| 17 | Kimi K2.6 | 2.38 | reproducedT1 | 2026-04-20 |
| 18 | Claude Sonnet 4.5 | 2.38 | reproducedT1 | 2025-09-29 |
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 18 models, with the best and worst results separated by 47.62 points. Because that spread is wide, the score differences likely reflect genuine separation between models rather than measurement noise.
3 cited facts
This benchmark records an integrity score of 99, an integrity grade of A, and an integrity rank of 6 among 61 benchmarks. The weakest component is saturation, meaning the ceiling is crowded and top scores become less decisive for comparing frontier models. These figures reflect the benchmark's health as a discriminator of frontier models under a disclosed harness, not any model's capability.
5 cited facts
This benchmark is not saturated, with the top score at 50.0 and a headroom of 50.0 points separating the leader from the ceiling. Only a single model clusters near the top, within a couple of points of the best result. Because it is not saturated, rank differences among the current top models reflect genuine capability separation rather than measurement noise.
5 cited facts
Task performance is reported under a consistent harness, so scores are directly comparable across runs using that harness. The privacy status of the test set is unknown, meaning if it is public, high scores would deserve more scrutiny for possible contamination than if it were held out. Contamination history is unknown, and the disclosed harness notes list metadata fields but do not characterize sensitivity to harness choices.
4 cited facts
This benchmark's intended scope is not documented: what it measures is unknown, so what it actually evaluates cannot be stated from the available facts. Its construction and key assumptions are likewise not documented, and no limitations are recorded; consequently the sharpest caveat is that any interpretation of results should be treated as provisional, since the benchmark provides no stated boundaries on what it can or cannot tell a reader.
4 cited facts
EBR-bench: EBR-bench as reported in Epoch AI's Capabilities Index CSV.
Claude Opus 5 leads EBR-bench at 50 — independently reproduced. The full leaderboard above lists every recorded measurement, not just the headline number.
18 models have recorded scores on EBR-bench, spanning a score spread of 47.62.
tensor.news grades EBR-bench A for integrity (score 99/100), ranking #6 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — EBR-bench still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on EBR-bench are reasonably apples-to-apples.
Every EBR-bench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.