DeepResearch Bench
Integrity rank #44 of 61 · 22 models scored · top score 55.31 · Claude Opus 4.6
unknown
Strongest on harness comparability, weakest on contamination resistance. 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 4.6 | 55.31 | unverifiedT1 | 2026-02-05 |
| 2 | GPT-5 | 55.13 | unverifiedT1 | 2025-08-07 |
| 3 | Claude Sonnet 4.6 | 54.87 | unverifiedT1 | 2026-02-17 |
| 4 | GPT-5.5 | 54.01 | unverifiedT1 | 2026-04-23 |
| 5 | Claude Sonnet 4.5 | 52.6 | unverifiedT1 | 2025-09-29 |
| 6 | Claude Opus 4.8 | 50.23 | unverifiedT1 | 2026-05-28 |
| 7 | Gemini 2.5 Pro (Jun 2025) | 49.71 | unverifiedT1 | 2025-06-05 |
| 8 | Claude Opus 4.1 | 49.7 | unverifiedT1 | 2025-08-05 |
| 9 | Claude Opus 4 | 49 | unverifiedT1 | 2025-05-22 |
| 10 | Grok 4 | 47.9 | unverifiedT1 | 2025-07-09 |
| 11 | Claude Sonnet 4 | 47.8 | unverifiedT1 | 2025-05-22 |
| 12 | o3 | 46.6 | unverifiedT1 | 2024-12-20 |
| 13 | Claude 3.7 Sonnet | 43.6 | unverifiedT1 | 2025-02-24 |
| 14 | GPT-5.1 | 42.79 | unverifiedT1 | 2025-11-13 |
| 15 | GPT-5.2 | 41.12 | unverifiedT1 | 2025-12-11 |
| 16 | Gemini 3.1 Flash-Lite | 36.39 | unverifiedT1 | 2026-03-03 |
| 17 | GPT-5.4 Mini | 36.26 | unverifiedT1 | 2026-03-17 |
| 18 | DeepSeek-R1 (May 2025) | 35.1 | unverifiedT1 | 2025-05-28 |
| 19 | GPT-5.4 | 35.09 | unverifiedT1 | 2026-03-05 |
| 20 | Gemini 2.5 Pro (May 2025) | 31.9 | unverifiedT2 | 2025-05-06 |
| 21 | GPT-4.1 | 29.31 | unverifiedT2 | 2025-04-14 |
| 22 | Gemini 2.5 Flash (Apr 2025) | 29.19 | unverifiedT2 | 2025-04-17 |
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.
Twenty-two models is middling depth; the ordering is informative, but new entrants can still visibly reshuffle it. The score spread is 26.12 points, meaning a wide gap supports real separation between models while a narrow gap suggests close scores may not reflect a genuine capability gap.
2 cited facts
What a B at 77 licenses is limited: ranked 36th of 51, this benchmark orders models usefully but sits far enough down the tracked set that close results deserve corroboration elsewhere. The single weakest component is contamination, which narrows how much to trust top scores because the test set privacy is unknown and the contamination window is wide. The score frames the benchmark's health as a discriminator of frontier models under a consistent harness, never any model's capability.
6 cited facts
The benchmark is not saturated, as documented. It has a top score of 55.31 with a score headroom of 44.69, leaving ample room to separate the field at the top. Four models cluster within a couple of points of the top, meaning small ranking differences are likely noise.
4 cited facts
Consistency plus disclosed reporting detail — model identities, performance, score provenance — lets a reader audit a gap instead of taking it on faith; that provenance trail matters more than the comparability label itself. The test set privacy status remains unknown, limiting assessment of potential contamination risks in reported performance. Given consistent harness comparability, low cross-harness score comparisons are not expected, and the unknown test set privacy requires scrutiny of performance claims for possible data leakage.
5 cited facts
What this benchmark measures is not documented. Construction, assumptions, and limitations are blank at once, which is as thin as benchmark documentation gets. The full burden of trust shifts to whoever cites the score.
4 cited facts
DeepResearch Bench: DeepResearch Bench as reported in Epoch AI's Capabilities Index CSV.
Claude Opus 4.6 leads DeepResearch Bench at 55.31. The full leaderboard above lists every recorded measurement, not just the headline number.
22 models have recorded scores on DeepResearch Bench, spanning a score spread of 26.12.
tensor.news grades DeepResearch Bench B for integrity (score 77/100), ranking #44 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — DeepResearch 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 DeepResearch Bench are reasonably apples-to-apples.
Every DeepResearch Bench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.