Benchmark Almanac

RECODE-H

GPT-5 84.3% leads since October 2025
Not in model cards

Agents implement research code from papers and repositories. They can request feedback from a simulated collaborator before their code runs against tests.

Made by
Chunyu Miao and colleagues
Released
October 2025
Kind
AI research tasks
Level
Expert
Scored
Checked automatically
Answers
Code
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code, Documents → Code, Actions, Text
Interaction
Agentic
Environment
Terminal, APIs and tools
Refresh
One fixed set
Run by
Anyone
In views
Agents
Strongest model tested
GPT-5 #71 on the capability index
Last model added
September 2025
Scores refreshed
November 20th, 2026
7
Entries in its paper, 6 models we track

GPT-5 leads with 84.3% among models released by September 2025 GPT-5 leads with 84.3% among models released by September 2025

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. 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 GPT-5 Feedback Level 4 84.3%
  2. 2 GPT-5 mini Feedback Level 4 81%
  3. 3 Gemini 2.5 Flash Sep 2025, Feedback Level 4 77.6%
  4. 4 DeepSeek-V3.1 Feedback Level 4 77.3%
  5. 5 Gemini-2.5-pro (snapshot unspecified) Feedback Level 4 70.6%
  6. 6 Claude Sonnet 4 Feedback Level 4 64.4%
  7. 7 GPT-5 nano Feedback Level 4 54.7%

Source: The RECODE-H paper

#EntryScoreSettingDateSource
1 GPT-5 84.3% Feedback Level 4, ReCodeAgent Aug 2025 Official leaderboard
2 GPT-5 mini 81% Feedback Level 4, ReCodeAgent Aug 2025 Official leaderboard
3 Gemini 2.5 Flash Sep 2025 77.6% Feedback Level 4, ReCodeAgent Sep 2025 Official leaderboard
4 DeepSeek-V3.1 77.3% Feedback Level 4, ReCodeAgent Aug 2025 Official leaderboard
5 Gemini-2.5-pro (snapshot unspecified) 70.6% Feedback Level 4, ReCodeAgent Official leaderboard
6 Claude Sonnet 4 64.4% Feedback Level 4, ReCodeAgent May 2025 Official leaderboard
7 GPT-5 nano 54.7% Feedback Level 4, ReCodeAgent Aug 2025 Official leaderboard
#ModelScoreSettingDateSource
1 GPT-5 84.3% Feedback Level 4, ReCodeAgent Aug 2025 Official leaderboard
2 GPT-5 mini 81% Feedback Level 4, ReCodeAgent Aug 2025 Official leaderboard
3 Gemini 2.5 Flash Sep 2025 77.6% Feedback Level 4, ReCodeAgent Sep 2025 Official leaderboard
4 DeepSeek-V3.1 77.3% Feedback Level 4, ReCodeAgent Aug 2025 Official leaderboard
5 Claude Sonnet 4 64.4% Feedback Level 4, ReCodeAgent May 2025 Official leaderboard
6 GPT-5 nano 54.7% Feedback Level 4, ReCodeAgent Aug 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.

Lowest effort to highest, light to dark
GPT-5 ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
84.3%
GPT-5 mini ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
81%
Gemini 2.5 Flash ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le… Feedback Le…
77.6%
DeepSeek-V3.1 ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
77.3%
Claude Sonnet 4 ReCodeAgent
Feedback Le… Feedback Le… Feedback Le…
64.4%
GPT-5 nano ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
54.7%

Source: The RECODE-H paper

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5 Feedback Level 4 ReCodeAgent 84.3%
GPT-5 Feedback Level 3 ReCodeAgent 83.7%
GPT-5 mini Feedback Level 4 ReCodeAgent 81%
Gemini 2.5 Flash Sep 2025 Feedback Level 4 ReCodeAgent 77.6%
GPT-5 Feedback Level 2 ReCodeAgent 77.4%
DeepSeek-V3.1 Feedback Level 4 ReCodeAgent 77.3%
GPT-5 mini Feedback Level 3 ReCodeAgent 73.2%
Gemini 2.5 Flash Sep 2025 Feedback Level 3 ReCodeAgent 70.7%
GPT-5 mini Feedback Level 2 ReCodeAgent 70.2%
GPT-5 Feedback Level 1 ReCodeAgent 68.8%
Claude Sonnet 4 Feedback Level 4 ReCodeAgent 64.4%
GPT-5 mini Feedback Level 1 ReCodeAgent 62.8%
DeepSeek-V3.1 Feedback Level 2 ReCodeAgent 60.6%
Gemini 2.5 Flash Sep 2025 Feedback Level 2 ReCodeAgent 58.7%
DeepSeek-V3.1 Feedback Level 3 ReCodeAgent 57.6%
GPT-5 nano Feedback Level 4 ReCodeAgent 54.7%
GPT-5 Feedback Level 0 ReCodeAgent 53.7%
Claude Sonnet 4 Feedback Level 3 ReCodeAgent 53.5%
Gemini 2.5 Flash Sep 2025 Feedback Level 1 ReCodeAgent 52.6%
Claude Sonnet 4 Feedback Level 2 ReCodeAgent 51.8%
Claude Sonnet 4 Feedback Level 1 ReCodeAgent 48.5%
DeepSeek-V3.1 Feedback Level 1 ReCodeAgent 47.2%
GPT-5 mini Feedback Level 0 ReCodeAgent 42.3%
GPT-5 nano Feedback Level 3 ReCodeAgent 42.2%
GPT-5 nano Feedback Level 2 ReCodeAgent 39.4%
Claude Sonnet 4 Feedback Level 0 ReCodeAgent 35.1%
Gemini 2.5 Flash Sep 2025 Feedback Level 0 ReCodeAgent 32.2%
GPT-5 nano Feedback Level 1 ReCodeAgent 31.9%
DeepSeek-V3.1 Feedback Level 0 ReCodeAgent 30.7%
GPT-5 nano Feedback Level 0 ReCodeAgent 23.8%

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.

Lowest effort to highest, light to dark
GPT-5 ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
84.3%
GPT-5 mini ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
81%
Gemini 2.5 Flash ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le… Feedback Le…
77.6%
DeepSeek-V3.1 ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
77.3%
Claude Sonnet 4 ReCodeAgent
Feedback Le… Feedback Le… Feedback Le…
64.4%
GPT-5 nano ReCodeAgent
Feedback Le… Feedback Le… Feedback Le… Feedback Le…
54.7%

Source: The RECODE-H paper

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5 Feedback Level 4 ReCodeAgent 84.3%
GPT-5 Feedback Level 3 ReCodeAgent 83.7%
GPT-5 mini Feedback Level 4 ReCodeAgent 81%
Gemini 2.5 Flash Sep 2025 Feedback Level 4 ReCodeAgent 77.6%
GPT-5 Feedback Level 2 ReCodeAgent 77.4%
DeepSeek-V3.1 Feedback Level 4 ReCodeAgent 77.3%
GPT-5 mini Feedback Level 3 ReCodeAgent 73.2%
Gemini 2.5 Flash Sep 2025 Feedback Level 3 ReCodeAgent 70.7%
GPT-5 mini Feedback Level 2 ReCodeAgent 70.2%
GPT-5 Feedback Level 1 ReCodeAgent 68.8%
Claude Sonnet 4 Feedback Level 4 ReCodeAgent 64.4%
GPT-5 mini Feedback Level 1 ReCodeAgent 62.8%
DeepSeek-V3.1 Feedback Level 2 ReCodeAgent 60.6%
Gemini 2.5 Flash Sep 2025 Feedback Level 2 ReCodeAgent 58.7%
DeepSeek-V3.1 Feedback Level 3 ReCodeAgent 57.6%
GPT-5 nano Feedback Level 4 ReCodeAgent 54.7%
GPT-5 Feedback Level 0 ReCodeAgent 53.7%
Claude Sonnet 4 Feedback Level 3 ReCodeAgent 53.5%
Gemini 2.5 Flash Sep 2025 Feedback Level 1 ReCodeAgent 52.6%
Claude Sonnet 4 Feedback Level 2 ReCodeAgent 51.8%
Claude Sonnet 4 Feedback Level 1 ReCodeAgent 48.5%
DeepSeek-V3.1 Feedback Level 1 ReCodeAgent 47.2%
GPT-5 mini Feedback Level 0 ReCodeAgent 42.3%
GPT-5 nano Feedback Level 3 ReCodeAgent 42.2%
GPT-5 nano Feedback Level 2 ReCodeAgent 39.4%
Claude Sonnet 4 Feedback Level 0 ReCodeAgent 35.1%
Gemini 2.5 Flash Sep 2025 Feedback Level 0 ReCodeAgent 32.2%
GPT-5 nano Feedback Level 1 ReCodeAgent 31.9%
DeepSeek-V3.1 Feedback Level 0 ReCodeAgent 30.7%
GPT-5 nano Feedback Level 0 ReCodeAgent 23.8%

Number one over time

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

OpenAI
Benchmark paper Not tested since Oct 2025
GPT-5 84.3%

Source: The RECODE-H paper

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
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) GPT-5 Claude Sonnet 4 0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) GPT-5 Claude Sonnet 4

Source: Official RECODE-H 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 benchmark’s paper (7 entries).
  • Pass Models scored 6 models we track, plus 1 other entry 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 GPT-5 (#71).
  • Pass Headroom The best is 84.3%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 30 rows for models we track and 1 other entry 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.

How we check