DeepResearch Bench
Integrity rank #54 of 65 · 24 models scored · top score 55.3 · Claude Opus 4.6
Multi-step deep-research agent capability on PhD-level research tasks across 22 academic domains.
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 | 55.3 | unverifiedT1 | 2026-02-05 |
| 2 | Claude Sonnet 4.6 | 54.9 | unverifiedT1 | 2026-02-17 |
| 3 | Claude Opus 4.5 | 54.8 | unverifiedT1 | 2025-11-24 |
| 4 | GPT-5.5 | 54 | unverifiedT1 | 2026-04-23 |
| 5 | Claude Sonnet 4.5 | 52.6 | unverifiedT1 | 2025-09-29 |
| 6 | Claude Opus 4.8 | 50.2 | unverifiedT1 | 2026-05-28 |
| 7 | Gemini 3 Flash | 49.8 | unverifiedT1 | 2025-12-17 |
| 8 | GPT-5 | 49.6 | unverifiedT1 | 2025-08-07 |
| 9 | Claude Opus 4.1 | 48.3 | unverifiedT1 | 2025-08-05 |
| 10 | Gemini 3.1 Pro | 47.8 | unverifiedT1 | 2026-02-19 |
| 11 | Grok 4 | 47.3 | unverifiedT1 | 2025-07-09 |
| 12 | Claude Opus 4 | 46.8 | unverifiedT1 | 2025-05-22 |
| 13 | Claude Sonnet 4 | 46.6 | unverifiedT1 | 2025-05-22 |
| 14 | Gemini 3 Pro | 46.3 | unverifiedT1 | 2025-11-18 |
| 15 | Claude Haiku 4.5 | 45.5 | unverifiedT1 | 2025-10-15 |
| 16 | o3 | 45.2 | unverifiedT1 | 2025-04-16 |
| 17 | Claude 3.7 Sonnet | 43.6 | unverifiedT1 | 2025-02-24 |
| 18 | Gemini 2.5 Pro (Jun 2025) | 42.8 | unverifiedT1 | 2025-06-05 |
| 19 | GPT-5.1 | 42.8 | unverifiedT1 | 2025-11-13 |
| 20 | GPT-5.2 | 41.1 | unverifiedT1 | 2025-12-11 |
| 21 | Gemini 3.1 Flash-Lite | 37.3 | unverifiedT1 | 2026-03-03 |
| 22 | GPT-5.4 Mini | 36.3 | unverifiedT1 | 2026-03-17 |
| 23 | DeepSeek-R1 (May 2025) | 35.1 | unverifiedT1 | 2025-05-28 |
| 24 | GPT-5.4 | 35.1 | unverifiedT1 | 2026-03-05 |
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 24 models, and the gap between its best and worst performers spans 20.2 points. A spread that wide supports real separation between models, so the rank differences observed here are more likely to reflect genuine capability gaps than to be an artifact of close scores.
3 cited facts
The benchmark's Benchmark Integrity Index is 71 (grade B), ranking 51st among 62 benchmarks. Its weakest integrity component is contamination, which narrows how much to trust top scores. This score reflects the benchmark's health as a discriminator of frontier models under a disclosed harness, not any model's capability.
6 cited facts
This benchmark is not saturated. The top score is 55.3, leaving 44.7 points of headroom to the ceiling, and four models cluster near the top within a couple of points. Because it is not saturated, there is still genuine room to separate the field at the top, rather than treating small ranking differences among the top cluster as noise.
6 cited facts
Scores for this benchmark are directly comparable under a consistent harness, and they should be read as task performance under the disclosed harness rather than as evidence of deployed capability; where comparability is low, scores from different harnesses should not be compared head-to-head. The test set privacy status is unknown, so it is not established whether the test set is public or held out; if public, high scores warrant more scrutiny for possible contamination than a held-out set would. Qualitatively, the contamination history notes public task prompts and web-search-enabled evaluation that may expose models to pre-seen material, so high scores should be interpreted with caution about contamination rather than as clean generalization.
4 cited facts
This benchmark measures multi-step deep-research agent capability on PhD-level research tasks across a range of academic domains. The test was constructed from expert-authored PhD-level tasks in English and Chinese, with domain distribution guided by analysis of real user queries from a search-enabled language model. Its design assumes that such tasks probe the upper limit of agent capability and that multi-step reasoning and comprehensive synthesis are the key discriminating axes, and it is calibrated against expert human report quality on the evaluation dimensions. The sharpest caveat is that the LLM-as-judge evaluation inherits LLM biases, so the benchmark cannot deliver a fully objective measure of agent capability.
5 cited facts
DeepResearch Bench: DeepResearch Bench as reported in Epoch AI's Capabilities Index CSV.
Claude Opus 4.6 leads DeepResearch Bench at 55.3. The full leaderboard above lists every recorded measurement, not just the headline number.
24 models have recorded scores on DeepResearch Bench, spanning a score spread of 20.2.
tensor.news grades DeepResearch Bench B for integrity (score 70/100), ranking #54 of 65 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 arxiv.org/abs/2506.11763.