ResearchCodeBench: leaderboard
ResearchCodeBench — unweighted first-completion task pass rate across 212 coding challenges from 20 machine-learning research papers; distinct from the website’s LoC-weighted score and the 109-task HARD subset.
Metric: Task Success Rate (%). Source: researchcodebench.github.io. Saturation forecast: Estimated already saturated. 32 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 05-06) | 64.15 | #147 |
| 2 | O3 (2025-04-16) (High) | 59.43 | #121 (O3 (2025-04-16)) |
| 3 | Gemini 2.5 Pro (Preview 03-25) | 58.96 | #123 |
| 4 | O4 Mini (High) | 58.49 | #180 (O4 Mini) |
| 5 | O3 Mini (2025-01-31) (High) | 52.36 | #240 (O3 Mini (2025-01-31)) |
| 6 | Claude 3.7 Sonnet (20250219) | 51.89 | #201 |
| 7 | GPT-4.1 | 50 | #248 |
| 8 | Claude 3.5 Sonnet (20241022) | 48.58 | #291 |
| 9 | O1 (2024-12-17) (High) | 48.11 | #149 (O1 (2024-12-17)) |
| 10 | DeepSeek R1 | 45.75 | #251 |
| 11 | Grok 3 | 42.92 | #304 |
| 12 | GPT-4.1 Mini | 42.45 | #355 |
| 13 | DeepSeek V3 (0324) | 42.45 | #340 |
| 14 | GPT-4o (2024-08-06) | 41.04 | #336 |
| 15 | Claude 3.5 Haiku | 37.74 | #568 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=researchcodebench · How It Works · Data refreshed daily, snapshot 2026-10-20.