RExBench
Agents extend 12 published AI research projects from expert instructions. The evaluator runs their code against task-specific success criteria. Harbor distributes two public tasks.
On November 19, 2026, Claude Opus 4.5 with OpenHands led individual runs at 50% final success. The leaderboard also lists GPT-5, o4-mini, and Claude Sonnet.
- Made by
- Nicholas Edwards and colleagues
- Released
- June 2025
- Kind
- AI research tasks
- Level
- Research frontier
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal
- Items from
- Real work and use, Written by experts
- Refresh
- One fixed set
- Run by
- A host
- In views
- Agents
- Strongest model tested
- Claude Opus 4.5 #62 on the capability index
- Last model added
- November 2025
- Scores refreshed
- November 20th, 2026
Claude Opus 4.5 leads with 50% among models released by November 2025
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
The newest model with a score here came out in November 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 1 Claude Opus 4.5 run2 50%
- 2 Claude Sonnet 4 run4 41.7%
- 3 Claude 3.7 Sonnet run3 33.3%
- 4 GPT-5 run1 33.3%
- 5 o4-mini run2 16.7%
- 6 o1 run1 0%
- 7
DeepSeek-R1 May 2025, run1 0%
Source: Official RExBench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.5 | 50% | run2, openhands | Nov 2025 | Official leaderboard |
| 2 | Claude Sonnet 4 | 41.7% | run4, openhands | May 2025 | Official leaderboard |
| 3 | Claude 3.7 Sonnet | 33.3% | run3, openhands | Feb 2025 | Official leaderboard |
| 4 | GPT-5 | 33.3% | run1, openhands | Aug 2025 | Official leaderboard |
| 5 | o4-mini | 16.7% | run2, openhands | Apr 2025 | Official leaderboard |
| 6 | o1 | 0% | run1, aider | Dec 2024 | Official leaderboard |
| 7 | | 0% | run1, aider | May 2025 | Official leaderboard |
Scores at each effort level
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: Official RExBench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Claude Opus 4.5 | run2 | openhands | 50% | ||
| Claude Sonnet 4 | run4 | openhands | 41.7% | ||
| Claude Opus 4.5 | run1 | openhands | 41.7% | ||
| Claude Opus 4.5 | run3 | openhands | 41.7% | ||
| Claude Opus 4.5 | run4 | openhands | 41.7% | ||
| Claude 3.7 Sonnet | run3 | openhands | 33.3% | ||
| Claude 3.7 Sonnet | run5 | openhands | 33.3% | ||
| Claude Sonnet 4 | run1 | openhands | 33.3% | ||
| Claude Sonnet 4 | run2 | openhands | 33.3% | ||
| Claude Sonnet 4 | run3 | openhands | 33.3% | ||
| Claude Sonnet 4 | run5 | openhands | 33.3% | ||
| GPT-5 | run1 | openhands | 33.3% | ||
| Claude Opus 4.5 | run5 | openhands | 33.3% | ||
| Claude 3.7 Sonnet | run1 | aider | 25% | ||
| Claude 3.7 Sonnet | run2 | openhands | 25% | ||
| Claude 3.7 Sonnet | run4 | openhands | 25% | ||
| GPT-5 | run2 | openhands | 25% | ||
| GPT-5 | run3 | openhands | 25% | ||
| GPT-5 | run4 | openhands | 25% | ||
| GPT-5 | run5 | openhands | 25% | ||
| Claude Opus 4.5 | run1 | aider | 25% | ||
| Claude Opus 4.5 | run3 | aider | 25% | ||
| Claude Opus 4.5 | run4 | aider | 25% | ||
| Claude Opus 4.5 | run5 | aider | 25% | ||
| Claude 3.7 Sonnet | run1 | openhands | 16.7% | ||
| o4-mini | run2 | openhands | 16.7% | ||
| o4-mini | run4 | openhands | 16.7% | ||
| Claude Sonnet 4 | run3 | aider | 16.7% | ||
| Claude Sonnet 4 | run4 | aider | 16.7% | ||
| Claude Sonnet 4 | run5 | aider | 16.7% | ||
| Claude Opus 4.5 | run2 | aider | 16.7% | ||
| Claude 3.7 Sonnet | run2 | aider | 8.33% | ||
| Claude 3.7 Sonnet | run3 | aider | 8.33% | ||
| o4-mini | run2 | aider | 8.33% | ||
| o4-mini | run4 | aider | 8.33% | ||
| o4-mini | run1 | openhands | 8.33% | ||
| Claude Sonnet 4 | run1 | aider | 8.33% | ||
| Claude Sonnet 4 | run2 | aider | 8.33% | ||
| GPT-5 | run1 | aider | 8.33% | ||
| o1 | run1 | aider | 0% | ||
| o1 | run2 | aider | 0% | ||
| o1 | run3 | aider | 0% | ||
| o1 | run4 | aider | 0% | ||
| o1 | run5 | aider | 0% | ||
| o1 | run1 | openhands | 0% | ||
| o1 | run2 | openhands | 0% | ||
| o1 | run3 | openhands | 0% | ||
| o1 | run4 | openhands | 0% | ||
| o1 | run5 | openhands | 0% | ||
| Claude 3.7 Sonnet | run4 | aider | 0% | ||
| Claude 3.7 Sonnet | run5 | aider | 0% | ||
| o4-mini | run1 | aider | 0% | ||
| o4-mini | run3 | aider | 0% | ||
| o4-mini | run5 | aider | 0% | ||
| o4-mini | run3 | openhands | 0% | ||
| o4-mini | run5 | openhands | 0% | ||
| run1 | aider | 0% | |||
| run2 | aider | 0% | |||
| run3 | aider | 0% | |||
| run4 | aider | 0% | |||
| run5 | aider | 0% | |||
| run1 | openhands | 0% | |||
| run2 | openhands | 0% | |||
| run3 | openhands | 0% | |||
| run4 | openhands | 0% | |||
| run5 | openhands | 0% | |||
| GPT-5 | run2 | aider | 0% | ||
| GPT-5 | run3 | aider | 0% | ||
| GPT-5 | run4 | aider | 0% | ||
| GPT-5 | run5 | aider | 0% |
2 models have held number one on the official leaderboard since its release in June 2025
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official RExBench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official RExBench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in AI research
- PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics.
Qwen 3.8 Max 93%
- MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
- PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
- MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6.2 Astra 96.1%
- PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Kimi K3 0.71
- KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup.
MiniMax-M3 28.8%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 12 months ago, in November 2025, for Claude Opus 4.5. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the official leaderboard (7 entries).
- Pass Models scored 7 models we track.
- Warning Newest score The newest came 12 months ago, in November 2025, for Claude Opus 4.5.
- Note Frontier models None of the top 10 has a score. The strongest tested is Claude Opus 4.5 (#62).
- Pass Headroom The best is 50%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on rexbench.com: 70 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.