DeepResearch Bench
PhD-level research reports with citations, graded for quality and citation accuracy.
Agent Index weight
5%
in the Knowledge-work agents pillar
Models tested
24
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 4.6 Supported Anthropic | 55.3% | source | — | — | |
| 2 | Claude Sonnet 4.6 Supported Anthropic | 54.9% | source | — | — | |
| 3 | Claude Opus 4.5 (latest) Supported Anthropic | 54.8% | source | — | — | |
| 4 | GPT-5.5 Supported OpenAI | 54.0% | source | — | — | |
| 5 | Claude Sonnet 4.5 (latest) Supported Anthropic | 52.6% | source | — | — | |
| 6 | Claude Opus 4.8 Supported Anthropic | 50.2% | source | — | — | |
| 7 | Gemini 3 Flash Preview Supported Google | 49.8% | source | — | — | |
| 8 | GPT-5 Supported OpenAI | 49.6% | source | — | — | |
| 9 | Claude Opus 4.1 (latest) Supported Anthropic | 48.3% | source | — | — | |
| 10 | Gemini 3.1 Pro Preview Supported Google | 47.8% | source | — | — | |
| 11 | Grok 4 Supported xAI | 47.3% | True | — | — | |
| 12 | Claude-Opus-4 Supported Anthropic | 46.8% | source | — | — | |
| 13 | Claude Sonnet 4 Supported Anthropic | 46.6% | source | — | — | |
| 14 | Gemini 3 Pro Preview Supported Google | 46.3% | source | — | — | |
| 15 | Claude Haiku 4.5 Estimated Anthropic | 45.5% | source | — | — | |
| 16 | o3 Supported OpenAI | 45.2% | source | — | — | |
| 17 | Claude 3.7 Sonnet Supported Anthropic | 43.6% | source | — | — | |
| 18 | Gemini 2.5 Pro Supported Google | 42.8% | source | — | — | |
| 19 | GPT-5.1 Supported OpenAI | 42.8% | source | — | — | |
| 20 | GPT-5.2 Supported OpenAI | 41.1% | source | — | — | |
| 21 | Gemini 3.1 Flash Lite Google | 37.3% | source | — | — | |
| 22 | GPT-5.4 mini OpenAI | 36.3% | source | — | — | |
| 23 | DeepSeek-R1 Estimated DeepSeek | 35.1% | source | — | — | |
| 24 | GPT-5.4 Supported OpenAI | 35.1% | source | — | — |
For models without a DeepResearch Bench result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 24 models (residual SD 0.24 on the logit scale). Estimates are labelled and widen the model’s range.