Benchmark Almanac

CS-4k

o3 71.3 leads since October 2025
Not in model cards

Models answer questions about the stages of computer science research using papers as evidence. A model judge scores answers on the held-out evaluation set.

Made by
Penghao Wang and colleagues
Released
October 2025
Kind
AI research tasks
Level
Expert
Scored
Graded by a model
Answers
Open-ended
Tasks
All public
Measures
Capability
Language
English
In and out
Text → Text
Interaction
Single turn
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
Strongest model tested
GPT-5 #71 on the capability index
Last model added
August 2025
Scores refreshed
November 20th, 2026
30
Entries in its paper, 18 models we track

o3 leads with 71.3 among models released by August 2025 o3 leads with 71.3 among models released by August 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. 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 August 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 o3 71.3
  2. 2 o4-mini 64.9
  3. 3 GPT-5 55.8
  4. 4 o3-mini 55.1
  5. 5 DeepSeek-R1 (snapshot unspecified) 54.7
  6. 6 Qwen2.5-7B-Instruct-GRPO-2RMs 53.1
  7. 7 Qwen2.5-7B-Instruct-GRPO 51.0
  8. 8 GPT-4.1 50.1
  9. 9 Gemini 2.5 Pro (snapshot unspecified) 49.5
  10. 10 Qwen3-235B-A22B-Thinking Jul 2025 49.0
  11. 11 GPT-4.1 mini 48.9
  12. 12 Gemini 2.5 Flash (snapshot unspecified) Search enabled 44.9
  13. 13 Qwen3-235B-A22B-Instruct Jul 2025 44.8
  14. 14 Qwen3-30B-A3B-Thinking Jul 2025 44.8
  15. 15 Grok 4 44.7
  16. 16 Qwen3-30B-A3B-Instruct Jul 2025 44.3
  17. 17 Qwen2.5-7B-Instruct-SFT 43.1
  18. 18 Grok-3 mini 43.0
  19. 19 Grok 3 mini fast 42.8
  20. 20 o1-mini 41.9
  21. 21 Claude Opus 4 40.7
  22. 22 Claude Sonnet 4 40.5
  23. 23 GPT-4o (snapshot unspecified) 40.4
  24. 24 GPT-4o-mini (snapshot unspecified) 39.1
  25. 25 Claude 3.7 Sonnet 37.8
  26. 26 Gemini 2.0 Flash Feb 2025 37.3
  27. 27 Llama 3.3 70B Instruct 37.0
  28. 28 Qwen2.5-7B-Instruct 35.8
  29. 29 GPT-4o mini Search Preview 27.3
  30. 30 Claude-3.5 (snapshot unspecified) 22.1

Source: The CS-4k paper

#EntryScoreSettingDateSource
1 o3 71.3 Apr 2025 Official leaderboard
2 o4-mini 64.9 Apr 2025 Official leaderboard
3 GPT-5 55.8 Aug 2025 Official leaderboard
4 o3-mini 55.1 Jan 2025 Official leaderboard
5 DeepSeek-R1 (snapshot unspecified) 54.7 Official leaderboard
6 Qwen2.5-7B-Instruct-GRPO-2RMs 53.1 Official leaderboard
7 Qwen2.5-7B-Instruct-GRPO 51.0 Official leaderboard
8 GPT-4.1 50.1 Apr 2025 Official leaderboard
9 Gemini 2.5 Pro (snapshot unspecified) 49.5 Official leaderboard
10 Qwen3-235B-A22B-Thinking Jul 2025 49.0 Jul 2025 Official leaderboard
11 GPT-4.1 mini 48.9 Apr 2025 Official leaderboard
12 Gemini 2.5 Flash (snapshot unspecified) 44.9 Search enabled Official leaderboard
13 Qwen3-235B-A22B-Instruct Jul 2025 44.8 Jul 2025 Official leaderboard
14 Qwen3-30B-A3B-Thinking Jul 2025 44.8 Jul 2025 Official leaderboard
15 Grok 4 44.7 Jul 2025 Official leaderboard
16 Qwen3-30B-A3B-Instruct Jul 2025 44.3 Jul 2025 Official leaderboard
17 Qwen2.5-7B-Instruct-SFT 43.1 Official leaderboard
18 Grok-3 mini 43.0 Apr 2025 Official leaderboard
19 Grok 3 mini fast 42.8 Official leaderboard
20 o1-mini 41.9 Sep 2024 Official leaderboard
21 Claude Opus 4 40.7 May 2025 Official leaderboard
22 Claude Sonnet 4 40.5 May 2025 Official leaderboard
23 GPT-4o (snapshot unspecified) 40.4 Official leaderboard
24 GPT-4o-mini (snapshot unspecified) 39.1 Official leaderboard
25 Claude 3.7 Sonnet 37.8 Feb 2025 Official leaderboard
26 Gemini 2.0 Flash Feb 2025 37.3 Feb 2025 Official leaderboard
27 Llama 3.3 70B Instruct 37.0 Dec 2024 Official leaderboard
28 Qwen2.5-7B-Instruct 35.8 Official leaderboard
29 GPT-4o mini Search Preview 27.3 Official leaderboard
30 Claude-3.5 (snapshot unspecified) 22.1 Official leaderboard
#ModelScoreSettingDateSource
1 o3 71.3 Apr 2025 Official leaderboard
2 o4-mini 64.9 Apr 2025 Official leaderboard
3 GPT-5 55.8 Aug 2025 Official leaderboard
4 o3-mini 55.1 Jan 2025 Official leaderboard
5 GPT-4.1 50.1 Apr 2025 Official leaderboard
6 Qwen3-235B-A22B-Thinking Jul 2025 49.0 Jul 2025 Official leaderboard
7 GPT-4.1 mini 48.9 Apr 2025 Official leaderboard
8 Qwen3-235B-A22B-Instruct Jul 2025 44.8 Jul 2025 Official leaderboard
9 Qwen3-30B-A3B-Thinking Jul 2025 44.8 Jul 2025 Official leaderboard
10 Grok 4 44.7 Jul 2025 Official leaderboard
11 Qwen3-30B-A3B-Instruct Jul 2025 44.3 Jul 2025 Official leaderboard
12 Grok-3 mini 43.0 Apr 2025 Official leaderboard
13 o1-mini 41.9 Sep 2024 Official leaderboard
14 Claude Opus 4 40.7 May 2025 Official leaderboard
15 Claude Sonnet 4 40.5 May 2025 Official leaderboard
16 Claude 3.7 Sonnet 37.8 Feb 2025 Official leaderboard
17 Gemini 2.0 Flash Feb 2025 37.3 Feb 2025 Official leaderboard
18 Llama 3.3 70B Instruct 37.0 Dec 2024 Official leaderboard

Rankings by part

We read 8 parts from the benchmark’s paper, each ranking models on one part of the test, such as Research domain, Previous methods and Existing challenges. o3 leads every one of them. Arrows show how far a model moved from its overall rank among the models we track.

Research domain

18 models we track have a score here.

  1. 1 o3 71.7
  2. 2 o4-mini 65.0
  3. 3 GPT-5 59.1
  4. 4 o3-mini 56.3
  5. 5 Qwen3-235B-A22B-Thinking Jul 2025 ↑1 51.6
  6. 6 GPT-4.1 ↓1 51.3
  7. 7 GPT-4.1 mini 49.7
  8. 8 Qwen3-235B-A22B-Instruct Jul 2025 46.5
  9. 9 Grok 4 ↑1 46.0
  10. 10 Claude Opus 4 ↑4 45.6

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
o3 71.3

Source: The CS-4k 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
20.0 40.0 60.0 80.0 2025 20.0 40.0 60.0 80.0 2025

Source: Official CS-4k 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

8 of 9 checks pass, 1 warning. Look first at the Newest score check. The newest came 15 months ago, in August 2025, for GPT-5. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 2 pass 2 notes
  • Pass Independent runs Scores from the benchmark’s paper (30 entries).
  • Pass Models scored 18 models we track, plus 12 other entries on its leaderboard.
  • Warning Newest score The newest came 15 months ago, in August 2025, for GPT-5.
  • Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5 (#71).
  • Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 19 rows for models we track and 12 other entries 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