AARRI-Bench
Agents complete research intern tasks that test scientific judgment, professional conduct, and attention to detail. The Harbor collection contains 82 tasks.
The June 2026 paper compares model and agent combinations. Claude Opus 4.7 with Mini-SWE-Agent led at 68.3% success.
- Made by
- AARR-Bench team
- Released
- April 2026
- Kind
- AI research tasks
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal
- Refresh
- One fixed set
- Run by
- Anyone
- Occupations
- Life, physical and social science
- In views
- Agents
- Strongest model tested
- Claude Opus 4.7 #31 on the capability index
- Last model added
- April 2026
- Scores refreshed
- November 20th, 2026
Claude Opus 4.7 leads with 68.3% among models released by April 2026
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.
The newest model with a score here came out in April 2026, 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 4.7 68.3%
- 2
Qwen 3.6 Plus 61.4%
- 3
DeepSeek-V4-Flash 60.5%
- 4
MiniMax-M2.7 58.1%
- 5 Kimi K2.6 56.4%
- 6 Claude Sonnet 4.6 54.4%
- 7 GPT-5.3 Codex 53.1%
Source: Official AARRI-Bench leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.7 | 68.3% | Mini-SWE-Agent | Apr 2026 | Official leaderboard |
| 2 | | 61.4% | Hermes Agent | Mar 2026 | Official leaderboard |
| 3 | | 60.5% | Mini-SWE-Agent | Apr 2026 | Official leaderboard |
| 4 | | 58.1% | Hermes Agent | Mar 2026 | Official leaderboard |
| 5 | Kimi K2.6 | 56.4% | Mini-SWE-Agent | Apr 2026 | Official leaderboard |
| 6 | Claude Sonnet 4.6 | 54.4% | Hermes Agent | Feb 2026 | Official leaderboard |
| 7 | GPT-5.3 Codex | 53.1% | Claude Code | Feb 2026 | Official leaderboard |
Rankings by part
We read 4 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Context, Mindset and Interaction. Claude Opus 4.7 leads 2 of them, and Qwen 3.6 Plus leads the other 2. Arrows show how far a model moved from its overall rank among the models we track.
Context
7 models we track have a score here.
- 1 Claude Opus 4.7 64.7%
- 2 Kimi K2.6 ↑3 59.4%
- 3
MiniMax-M2.7 ↑1 55.9%
- 4
DeepSeek-V4-Flash ↓1 55.9%
- 5 Claude Sonnet 4.6 ↑1 50%
- 6
Qwen 3.6 Plus ↓4 50%
- 7 GPT-5.3 Codex 47.1%
Mindset
7 models we track have a score here.
- 1
Qwen 3.6 Plus ↑1 76.9%
- 2 Claude Opus 4.7 ↓1 76.9%
- 3
DeepSeek-V4-Flash 76.9%
- 4
MiniMax-M2.7 69.2%
- 5 Claude Sonnet 4.6 ↑1 61.5%
- 6 Kimi K2.6 ↓1 61.5%
- 7 GPT-5.3 Codex 53.8%
Interaction
7 models we track have a score here.
- 1 Claude Opus 4.7 76.2%
- 2
DeepSeek-V4-Flash ↑1 76.2%
- 3
Qwen 3.6 Plus ↓1 71.4%
- 4 Claude Sonnet 4.6 ↑2 66.7%
- 5
MiniMax-M2.7 ↓1 66.7%
- 6 GPT-5.3 Codex ↑1 65%
- 7 Kimi K2.6 ↓2 65%
Hands-on
7 models we track have a score here.
- 1
Qwen 3.6 Plus ↑1 64.3%
- 2
MiniMax-M2.7 ↑2 57.1%
- 3 Claude Opus 4.7 ↓2 57.1%
- 4 GPT-5.3 Codex ↑3 50%
- 5 Claude Sonnet 4.6 ↑1 50%
- 6 Kimi K2.6 ↓1 50%
- 7
DeepSeek-V4-Flash ↓4 50%
2 models have held number one on the official leaderboard since its release in April 2026
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official AARRI-Bench 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 AARRI-Bench 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 7 months ago, in April 2026, for DeepSeek-V4-Flash. 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 7 months ago, in April 2026, for DeepSeek-V4-Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is Claude Opus 4.7 (#31).
- Pass Headroom The best is 68.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on aarr-bench.com: 16 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.