Benchmark Almanac

ResearchCodeBench

Gemini 2.5 Pro Jun 2025 64.2% leads since June 2025
Not in model cards

Agents read machine learning papers and fill missing Python code in the accompanying implementations. Unit tests check each of 212 snippets from 20 papers.

The comparison available November 19, 2026 tests GPT-4o mini and GPT-4.1 mini with Codex. GPT-4.1 mini led at 41.01% originally and 40.74% on Harbor.

Released
May 2025
Kind
Machine learning engineering
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code, Documents → Code, Actions
Interaction
Agentic
Environment
Terminal
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
In views
Agents
Strongest model tested
o3 #84 on the capability index
Last model added
September 2025
Scores refreshed
November 21st, 2026
31
Entries on its own leaderboard, 22 models we track

Gemini 2.5 Pro leads with 64.2% among models released by September 2025 Gemini 2.5 Pro leads with 64.2% among models released by September 2025

Best score per entry from the benchmark's own leaderboard, read on November 21st, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.

The newest model with a score here came out in September 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 Gemini 2.5 Pro Jun 2025 64.2%
  2. 2 o3 high 59.4%
  3. 3 o4-mini high 58.5%
  4. 4 o3-mini high 52.4%
  5. 5 Claude 3.7 Sonnet 51.9%
  6. 6 GPT-4.1 50%
  7. 7 Claude-3.5-Sonnet (snapshot unspecified) 48.6%
  8. 8 o1 high 48.1%
  9. 9 DeepSeek-R1 (snapshot unspecified) 45.8%
  10. 10 Gemini 2.5 Flash Sep 2025 45.3%
  11. 11 Grok 3 42.9%
  12. 12 DeepSeek-Chat-V3-0324 42.5%
  13. 13 GPT-4.1 mini 42.5%
  14. 14 GPT-4o (snapshot unspecified) 41%
  15. 15 Claude 3.5 Haiku 37.7%
  16. 16 Gemini 2.0 Flash Feb 2025 37.3%
  17. 17 Mistral Medium 3 35.4%
  18. 18 Mistral-Codestral-2501 33.5%
  19. 19 Command-A 31.1%
  20. 20 Gemini 2.0 Flash-Lite 30.7%
  21. 21 Llama 4 Maverick 27.4%
  22. 22 Qwen2.5-Coder-32B 25.9%
  23. 23 Nova-Pro-V1 25%
  24. 24 GPT-4o-Mini (snapshot unspecified) 23.1%
  25. 25 Grok-2 Dec 2024 22.2%
  26. 26 Grok-3 mini high 19.8%
  27. 27 Llama 4 Scout 18.4%
  28. 28 GPT-4.1 nano 15.1%
  29. 29 Llama 3.3 70B 12.3%
  30. 30 Qwen2.5-Turbo 8.02%
  31. 31 Mistral-Codestral-Mamba 1.42%

Source: Official ResearchCodeBench leaderboard

#EntryScoreSettingDateSource
1 Gemini 2.5 Pro Jun 2025 64.2% Jun 2025 Official leaderboard
2 o3 59.4% high Apr 2025 Official leaderboard
3 o4-mini 58.5% high Apr 2025 Official leaderboard
4 o3-mini 52.4% high Jan 2025 Official leaderboard
5 Claude 3.7 Sonnet 51.9% Feb 2025 Official leaderboard
6 GPT-4.1 50% Apr 2025 Official leaderboard
7 Claude-3.5-Sonnet (snapshot unspecified) 48.6% Official leaderboard
8 o1 48.1% high Dec 2024 Official leaderboard
9 DeepSeek-R1 (snapshot unspecified) 45.8% Official leaderboard
10 Gemini 2.5 Flash Sep 2025 45.3% Sep 2025 Official leaderboard
11 Grok 3 42.9% Apr 2025 Official leaderboard
12 DeepSeek-Chat-V3-0324 42.5% Official leaderboard
13 GPT-4.1 mini 42.5% Apr 2025 Official leaderboard
14 GPT-4o (snapshot unspecified) 41% Official leaderboard
15 Claude 3.5 Haiku 37.7% Oct 2024 Official leaderboard
16 Gemini 2.0 Flash Feb 2025 37.3% Feb 2025 Official leaderboard
17 Mistral Medium 3 35.4% May 2025 Official leaderboard
18 Mistral-Codestral-2501 33.5% Official leaderboard
19 Command-A 31.1% Official leaderboard
20 Gemini 2.0 Flash-Lite 30.7% Feb 2025 Official leaderboard
21 Llama 4 Maverick 27.4% Apr 2025 Official leaderboard
22 Qwen2.5-Coder-32B 25.9% Sep 2024 Official leaderboard
23 Nova-Pro-V1 25% Official leaderboard
24 GPT-4o-Mini (snapshot unspecified) 23.1% Official leaderboard
25 Grok-2 Dec 2024 22.2% Dec 2024 Official leaderboard
26 Grok-3 mini 19.8% high Apr 2025 Official leaderboard
27 Llama 4 Scout 18.4% Apr 2025 Official leaderboard
28 GPT-4.1 nano 15.1% Apr 2025 Official leaderboard
29 Llama 3.3 70B 12.3% Dec 2024 Official leaderboard
30 Qwen2.5-Turbo 8.02% Nov 2024 Official leaderboard
31 Mistral-Codestral-Mamba 1.42% Official leaderboard
#ModelScoreSettingDateSource
1 Gemini 2.5 Pro Jun 2025 64.2% Jun 2025 Official leaderboard
2 o3 59.4% high Apr 2025 Official leaderboard
3 o4-mini 58.5% high Apr 2025 Official leaderboard
4 o3-mini 52.4% high Jan 2025 Official leaderboard
5 Claude 3.7 Sonnet 51.9% Feb 2025 Official leaderboard
6 GPT-4.1 50% Apr 2025 Official leaderboard
7 o1 48.1% high Dec 2024 Official leaderboard
8 Gemini 2.5 Flash Sep 2025 45.3% Sep 2025 Official leaderboard
9 Grok 3 42.9% Apr 2025 Official leaderboard
10 GPT-4.1 mini 42.5% Apr 2025 Official leaderboard
11 Claude 3.5 Haiku 37.7% Oct 2024 Official leaderboard
12 Gemini 2.0 Flash Feb 2025 37.3% Feb 2025 Official leaderboard
13 Mistral Medium 3 35.4% May 2025 Official leaderboard
14 Gemini 2.0 Flash-Lite 30.7% Feb 2025 Official leaderboard
15 Llama 4 Maverick 27.4% Apr 2025 Official leaderboard
16 Qwen2.5-Coder-32B 25.9% Sep 2024 Official leaderboard
17 Grok-2 Dec 2024 22.2% Dec 2024 Official leaderboard
18 Grok-3 mini 19.8% high Apr 2025 Official leaderboard
19 Llama 4 Scout 18.4% Apr 2025 Official leaderboard
20 GPT-4.1 nano 15.1% Apr 2025 Official leaderboard
21 Llama 3.3 70B 12.3% Dec 2024 Official leaderboard
22 Qwen2.5-Turbo 8.02% Nov 2024 Official leaderboard

2 models have held number one on the official leaderboard since its release in May 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI Google DeepMind
Own leaderboard Not tested since Sep 2025

Source: Official ResearchCodeBench leaderboard

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) Gemini 2.5 Pro o3 o3-mini o1 Claude 3.5 Haiku Qwen2.5-Coder-32B 0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) Gemini 2.5 Pro o3-mini

Source: Official ResearchCodeBench 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

All 164 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 14 months ago, in September 2025, for Gemini 2.5 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 (31 entries).
  • Pass Models scored 22 models we track, plus 9 other entries on its leaderboard.
  • Warning Newest score The newest came 14 months ago, in September 2025, for Gemini 2.5 Flash.
  • Note Frontier models None of the top 10 has a score. The strongest tested is o3 (#84).
  • Pass Headroom The best is 64.2%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on github.com: 23 rows for models we track and 9 other entries at the last read.
  • Pass Last read November 21st, 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.

How we check