Benchmark Almanac

K-MetBench

Gemini 3 Pro 93.7% leads since April 2026
Not in model cards

Models answer 1,774 questions from the Korean National Meteorological Engineer Examination under a zero-shot protocol. The benchmark reports answer accuracy across five subjects and separate visual and Korean-specific subsets. Expert rationales support an additional rubric score for reasoning.

The paper evaluates 55 models. The opened leaderboard data contains 71 rows, including separate thinking settings. The 82 visual questions, 141 reasoning questions, and 73 Korean-specific questions are overlapping diagnostic subsets.

Released
April 2026
Kind
Science questions
Level
Expert
Scored
Checked automatically
Answers
Multiple choice
Tasks
All public
Measures
Capability
Language
Korean
In and out
Text, Images → Text
Interaction
Single turn
Items from
Exams and contests
Refresh
One fixed set
Run by
Anyone
Occupations
Architecture and engineering
In views
Vision
Strongest model tested
Kimi K3 #25 on the capability index
Last model added
July 2026
Scores refreshed
November 20th, 2026

In the authors’ words

The development of practical (multimodal) large language model assistants for Korean weather forecasters is hindered by the absence of a multidimensional, expert-level evaluation framework grounded in authoritative sources. To address this, we introduce K-MetBench, a diagnostic benchmark grounded in national qualification exams. It exposes critical gaps across four dimensions: expert visual reasoning of charts, logical validity via expert-verified rationales, Korean-specific geo-cultural comprehension, and fine-grained domain analysis. Our evaluation of 55 models reveals a profound modality gap in interpreting specialized diagrams and a reasoning gap where models hallucinate logic despite correct predictions. Crucially, Korean models outperform significantly larger global models in local contexts, demonstrating that parameter scaling alone cannot resolve cultural dependencies. K-MetBench serves as a roadmap for developing reliable, culturally aware expert AI agents. The dataset is available at https://huggingface.co/datasets/soyeonbot/K-MetBench .

Abstract of K-MetBench: A Multi-Dimensional Benchmark for Fine-Grained Evaluation of Expert Reasoning, Locality, and Multimodality in Meteorology (arXiv:2604.24645, CC0)
65
Entries on its own leaderboard, 43 models we track

Gemini 3 Pro leads with 93.7% among models released by July 2026 Gemini 3 Pro leads with 93.7% among models released by July 2026

Best score per entry 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. 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 July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

65 entries. Scroll sideways for the rest.

  1. 93.7 Gemini 3 Pro
  2. 87.8 GPT-5.2
  3. 86.5 Gemma 4 31B IT
  4. 84.4 Qwen3-VL-235B-A22B
  5. 83.0 Qwen3.5-27B
  6. 82.9 Qwen 3.6 35B-A3B
  7. 82.6 DeepSeek-V4-Pro
  8. 82.5 Kimi K2.6
  9. 81.8 Qwen3.5-35B-A3B
  10. 81.5 Kimi K3
  11. 80.1 DeepSeek V4 Flash 0731
  12. 79.6 DeepSeek-V4-Flash
  13. 78.6 Qwen3-VL-32B-Thinking
  14. 77.8 Command A Reasoning
  15. 77.3 gpt-oss-120b
  16. 76.7 Qwen3-30B-A3B-Thinking
  17. 76.1 A.X-4.0
  18. 75.9 EXAONE 4.5 33B
  19. 74.9 Qwen3-VL-30B-A3B-Thinking
  20. 74.9 Qwen3.5-9B
  21. 73.7 Qwen3-14B
  22. 72.4 Qwen3-VL-235B-A22B-Instruct
  23. 71.7 Qwen3-VL-8B-Thinking
  24. 71.5 gpt-oss-20b
  25. 70.1 Qwen3-8B
  26. 67.8 Qwen3-4B-Thinking-2507
  27. 67.5 Qwen3-VL-32B-Instruct
  28. 67.1 Qwen2.5-VL-72B
  29. 66.1 Qwen3-VL-4B-Thinking
  30. 65.5 c4ai-command-a-03-2025
  31. 64.7 Qwen3-30B-A3B-Instruct
  32. 62.2 Qwen3-VL-30B-A3B-Instruct
  33. 60.1 Qwen2.5-VL-32B
  34. 59.9 Llama-3.1-70B-Instruct
  35. 59.9 EXAONE 4.0 32B
  36. 58.7 VARCO-Vision-2.0-14B
  37. 57.3 InternVL3.5-38B-Instruct
  38. 56.9 Llama 3.2 90B
  39. 55.7 A.X-4.0-Light
  40. 53.8 Qwen3-VL-8B-Instruct
  41. 52.5 A.X-4.0-VL-Light
  42. 51.5 qwen3-4b-instruct-2507
  43. 51.5 Phi-4
  44. 51.0 Qwen3-VL-4B-Instruct
  45. 50.8 HyperCLOVAX-SEED-Think-14B
  46. 47.9 InternVL3.5-14B-Instruct
  47. 47.5 Qwen3-32B
  48. 46.8 Qwen3-1.7B
  49. 46.1 InternVL3.5-8B-Instruct
  50. 46.0 Qwen2.5-VL-7B
  51. 41.8 Llama-3.1-8B-Instruct
  52. 41.5 InternVL3.5-4B-Instruct
  53. 40.9 Qwen2.5-VL-3B
  54. 37.4 EXAONE-4.0-1.2B
  55. 35.2 VARCO-Vision-2.0-1.7B
  56. 33.8 Llama-3.2-3B-Instruct
  57. 32.1 Qwen3-0.6B
  58. 32.0 HyperCLOVAX-SEED-Vision-Instruct-3B
  59. 31.0 InternVL3.5-2B-Instruct
  60. 30.6 HyperCLOVAX-SEED-Text-Instruct-1.5B
  61. 30.4 Phi-4-mini-instruct
  62. 23.8 InternVL3.5-1B-Instruct
  63. 13.2 HyperCLOVAX-SEED-Text-Instruct-0.5B
  64. 12.6 Phi-4-mini-reasoning
  65. 3.55 Llama-3.2-1B-Instruct
  1. 1 Gemini 3 Pro thinking 93.7%
  2. 2 GPT-5.2 thinking 87.8%
  3. 3 Gemma 4 31B IT thinking 86.5%
  4. 4 Qwen3-VL-235B-A22B thinking 84.4%
  5. 5 Qwen3.5-27B thinking 83%
  6. 6 Qwen 3.6 35B-A3B thinking 82.9%
  7. 7 DeepSeek-V4-Pro 82.6%
  8. 8 Kimi K2.6 82.5%
  9. 9 Qwen3.5-35B-A3B thinking 81.8%
  10. 10 Kimi K3 81.5%
  11. 11 DeepSeek V4 Flash 0731 80.1%
  12. 12 DeepSeek-V4-Flash 79.6%
  13. 13 Qwen3-VL-32B-Thinking 78.6%
  14. 14 Command A Reasoning 77.8%
  15. 15 gpt-oss-120b 77.3%
  16. 16 Qwen3-30B-A3B-Thinking Jul 2025 76.7%
  17. 17 A.X-4.0 76.1%
  18. 18 EXAONE 4.5 33B 75.9%
  19. 19 Qwen3-VL-30B-A3B-Thinking 74.9%
  20. 20 Qwen3.5-9B thinking 74.9%
  21. 21 Qwen3-14B 73.7%
  22. 22 Qwen3-VL-235B-A22B-Instruct 72.4%
  23. 23 Qwen3-VL-8B-Thinking 71.7%
  24. 24 gpt-oss-20b 71.5%
  25. 25 Qwen3-8B 70.1%
  26. 26 Qwen3-4B-Thinking-2507 67.8%
  27. 27 Qwen3-VL-32B-Instruct 67.5%
  28. 28 Qwen2.5-VL-72B 67.1%
  29. 29 Qwen3-VL-4B-Thinking 66.1%
  30. 30 c4ai-command-a-03-2025 65.5%
  31. 31 Qwen3-30B-A3B-Instruct Jul 2025 64.7%
  32. 32 Qwen3-VL-30B-A3B-Instruct 62.2%
  33. 33 Qwen2.5-VL-32B 60.1%
  34. 34 Llama-3.1-70B-Instruct 59.9%
  35. 35 EXAONE 4.0 32B 59.9%
  36. 36 VARCO-Vision-2.0-14B 58.7%
  37. 37 InternVL3.5-38B-Instruct 57.3%
  38. 38 Llama 3.2 90B 56.9%
  39. 39 A.X-4.0-Light 55.7%
  40. 40 Qwen3-VL-8B-Instruct 53.8%
  41. 41 A.X-4.0-VL-Light 52.5%
  42. 42 qwen3-4b-instruct-2507 51.5%
  43. 43 Phi-4 51.5%
  44. 44 Qwen3-VL-4B-Instruct 51%
  45. 45 HyperCLOVAX-SEED-Think-14B 50.8%
  46. 46 InternVL3.5-14B-Instruct 47.9%
  47. 47 Qwen3-32B 47.5%
  48. 48 Qwen3-1.7B 46.8%
  49. 49 InternVL3.5-8B-Instruct 46.1%
  50. 50 Qwen2.5-VL-7B 46%
  51. 51 Llama-3.1-8B-Instruct 41.8%
  52. 52 InternVL3.5-4B-Instruct 41.5%
  53. 53 Qwen2.5-VL-3B 40.9%
  54. 54 EXAONE-4.0-1.2B 37.4%
  55. 55 VARCO-Vision-2.0-1.7B 35.2%
  56. 56 Llama-3.2-3B-Instruct 33.8%
  57. 57 Qwen3-0.6B 32.1%
  58. 58 HyperCLOVAX-SEED-Vision-Instruct-3B 32%
  59. 59 InternVL3.5-2B-Instruct 31%
  60. 60 HyperCLOVAX-SEED-Text-Instruct-1.5B 30.6%
  61. 61 Phi-4-mini-instruct 30.4%
  62. 62 InternVL3.5-1B-Instruct 23.8%
  63. 63 HyperCLOVAX-SEED-Text-Instruct-0.5B 13.2%
  64. 64 Phi-4-mini-reasoning 12.6%
  65. 65 Llama-3.2-1B-Instruct 3.55%

Source: Official K-MetBench leaderboard

Source: Official K-MetBench leaderboard

#EntryScoreSettingReleasedRun onSource
1 Gemini 3 Pro 93.7% thinking Nov 2025 Dec 2025 Official leaderboard
2 GPT-5.2 87.8% thinking Dec 2025 Dec 2025 Official leaderboard
3 Gemma 4 31B IT 86.5% thinking Apr 2026 Jun 2026 Official leaderboard
4 Qwen3-VL-235B-A22B 84.4% thinking Sep 2025 Oct 2025 Official leaderboard
5 Qwen3.5-27B 83% thinking Feb 2026 May 2026 Official leaderboard
6 Qwen 3.6 35B-A3B 82.9% thinking Apr 2026 May 2026 Official leaderboard
7 DeepSeek-V4-Pro 82.6% Apr 2026 Aug 2026 Official leaderboard
8 Kimi K2.6 82.5% Apr 2026 Aug 2026 Official leaderboard
9 Qwen3.5-35B-A3B 81.8% thinking Feb 2026 May 2026 Official leaderboard
10 Kimi K3 81.5% Jul 2026 Aug 2026 Official leaderboard
11 DeepSeek V4 Flash 0731 80.1% Jul 2026 Aug 2026 Official leaderboard
12 DeepSeek-V4-Flash 79.6% Apr 2026 Aug 2026 Official leaderboard
13 Qwen3-VL-32B-Thinking 78.6% Oct 2025 Oct 2025 Official leaderboard
14 Command A Reasoning 77.8% Aug 2025 Oct 2025 Official leaderboard
15 gpt-oss-120b 77.3% Aug 2025 Oct 2025 Official leaderboard
16 Qwen3-30B-A3B-Thinking Jul 2025 76.7% Jul 2025 Dec 2025 Official leaderboard
17 A.X-4.0 72B 76.1% Oct 2025 Official leaderboard
18 EXAONE 4.5 33B 75.9% Apr 2026 Apr 2026 Official leaderboard
19 Qwen3-VL-30B-A3B-Thinking 74.9% Oct 2025 Dec 2025 Official leaderboard
20 Qwen3.5-9B 74.9% thinking Feb 2026 May 2026 Official leaderboard
21 Qwen3-14B 73.7% Apr 2025 Oct 2025 Official leaderboard
22 Qwen3-VL-235B-A22B-Instruct 72.4% Sep 2025 Oct 2025 Official leaderboard
23 Qwen3-VL-8B-Thinking 71.7% Oct 2025 Oct 2025 Official leaderboard
24 gpt-oss-20b 71.5% Aug 2025 Oct 2025 Official leaderboard
25 Qwen3-8B 70.1% Apr 2025 Oct 2025 Official leaderboard
26 Qwen3-4B-Thinking-2507 67.8% Aug 2025 Oct 2025 Official leaderboard
27 Qwen3-VL-32B-Instruct 67.5% Oct 2025 Oct 2025 Official leaderboard
28 Qwen2.5-VL-72B 67.1% Jan 2025 Dec 2025 Official leaderboard
29 Qwen3-VL-4B-Thinking 66.1% Oct 2025 Oct 2025 Official leaderboard
30 c4ai-command-a-03-2025 111B 65.5% Oct 2025 Official leaderboard
31 Qwen3-30B-A3B-Instruct Jul 2025 64.7% Jul 2025 Oct 2025 Official leaderboard
32 Qwen3-VL-30B-A3B-Instruct 62.2% Oct 2025 Oct 2025 Official leaderboard
33 Qwen2.5-VL-32B 60.1% Mar 2025 Dec 2025 Official leaderboard
34 Llama-3.1-70B-Instruct 71B 59.9% Dec 2025 Official leaderboard
35 EXAONE 4.0 32B 59.9% Jul 2025 Oct 2025 Official leaderboard
36 VARCO-Vision-2.0-14B 14B 58.7% Oct 2025 Official leaderboard
37 InternVL3.5-38B-Instruct 38B 57.3% Oct 2025 Official leaderboard
38 Llama 3.2 90B 56.9% Sep 2024 Oct 2025 Official leaderboard
39 A.X-4.0-Light 7B 55.7% Oct 2025 Official leaderboard
40 Qwen3-VL-8B-Instruct 53.8% Oct 2025 Oct 2025 Official leaderboard
41 A.X-4.0-VL-Light 8B 52.5% Oct 2025 Official leaderboard
42 qwen3-4b-instruct-2507 51.5% Aug 2025 Oct 2025 Official leaderboard
43 Phi-4 51.5% Dec 2024 Dec 2025 Official leaderboard
44 Qwen3-VL-4B-Instruct 51% Oct 2025 Oct 2025 Official leaderboard
45 HyperCLOVAX-SEED-Think-14B 15B 50.8% Dec 2025 Official leaderboard
46 InternVL3.5-14B-Instruct 15B 47.9% Oct 2025 Official leaderboard
47 Qwen3-32B 47.5% Apr 2025 Oct 2025 Official leaderboard
48 Qwen3-1.7B 46.8% Apr 2025 Oct 2025 Official leaderboard
49 InternVL3.5-8B-Instruct 8B 46.1% Oct 2025 Official leaderboard
50 Qwen2.5-VL-7B 46% Jan 2025 Dec 2025 Official leaderboard
51 Llama-3.1-8B-Instruct 8B 41.8% Oct 2025 Official leaderboard
52 InternVL3.5-4B-Instruct 5B 41.5% Oct 2025 Official leaderboard
53 Qwen2.5-VL-3B 40.9% Jan 2025 Dec 2025 Official leaderboard
54 EXAONE-4.0-1.2B 1.2B 37.4% Oct 2025 Official leaderboard
55 VARCO-Vision-2.0-1.7B 1.7B 35.2% Oct 2025 Official leaderboard
56 Llama-3.2-3B-Instruct 3B 33.8% Oct 2025 Official leaderboard
57 Qwen3-0.6B 32.1% Apr 2025 Oct 2025 Official leaderboard
58 HyperCLOVAX-SEED-Vision-Instruct-3B 3B 32% Oct 2025 Official leaderboard
59 InternVL3.5-2B-Instruct 2B 31% Oct 2025 Official leaderboard
60 HyperCLOVAX-SEED-Text-Instruct-1.5B 1.5B 30.6% Dec 2025 Official leaderboard
61 Phi-4-mini-instruct 4B 30.4% Oct 2025 Official leaderboard
62 InternVL3.5-1B-Instruct 1B 23.8% Oct 2025 Official leaderboard
63 HyperCLOVAX-SEED-Text-Instruct-0.5B 0.5B 13.2% Oct 2025 Official leaderboard
64 Phi-4-mini-reasoning 12.6% Apr 2025 Dec 2025 Official leaderboard
65 Llama-3.2-1B-Instruct 1B 3.55% Dec 2025 Official leaderboard
#ModelScoreSettingReleasedRun onSource
1 Gemini 3 Pro 93.7% thinking Nov 2025 Dec 2025 Official leaderboard
2 GPT-5.2 87.8% thinking Dec 2025 Dec 2025 Official leaderboard
3 Gemma 4 31B IT 86.5% thinking Apr 2026 Jun 2026 Official leaderboard
4 Qwen3-VL-235B-A22B 84.4% thinking Sep 2025 Oct 2025 Official leaderboard
5 Qwen3.5-27B 83% thinking Feb 2026 May 2026 Official leaderboard
6 Qwen 3.6 35B-A3B 82.9% thinking Apr 2026 May 2026 Official leaderboard
7 DeepSeek-V4-Pro 82.6% Apr 2026 Aug 2026 Official leaderboard
8 Kimi K2.6 82.5% Apr 2026 Aug 2026 Official leaderboard
9 Qwen3.5-35B-A3B 81.8% thinking Feb 2026 May 2026 Official leaderboard
10 Kimi K3 81.5% Jul 2026 Aug 2026 Official leaderboard
11 DeepSeek V4 Flash 0731 80.1% Jul 2026 Aug 2026 Official leaderboard
12 DeepSeek-V4-Flash 79.6% Apr 2026 Aug 2026 Official leaderboard
13 Qwen3-VL-32B-Thinking 78.6% Oct 2025 Oct 2025 Official leaderboard
14 Command A Reasoning 77.8% Aug 2025 Oct 2025 Official leaderboard
15 gpt-oss-120b 77.3% Aug 2025 Oct 2025 Official leaderboard
16 Qwen3-30B-A3B-Thinking Jul 2025 76.7% Jul 2025 Dec 2025 Official leaderboard
17 EXAONE 4.5 33B 75.9% Apr 2026 Apr 2026 Official leaderboard
18 Qwen3-VL-30B-A3B-Thinking 74.9% Oct 2025 Dec 2025 Official leaderboard
19 Qwen3.5-9B 74.9% thinking Feb 2026 May 2026 Official leaderboard
20 Qwen3-14B 73.7% Apr 2025 Oct 2025 Official leaderboard
21 Qwen3-VL-235B-A22B-Instruct 72.4% Sep 2025 Oct 2025 Official leaderboard
22 Qwen3-VL-8B-Thinking 71.7% Oct 2025 Oct 2025 Official leaderboard
23 gpt-oss-20b 71.5% Aug 2025 Oct 2025 Official leaderboard
24 Qwen3-8B 70.1% Apr 2025 Oct 2025 Official leaderboard
25 Qwen3-4B-Thinking-2507 67.8% Aug 2025 Oct 2025 Official leaderboard
26 Qwen3-VL-32B-Instruct 67.5% Oct 2025 Oct 2025 Official leaderboard
27 Qwen2.5-VL-72B 67.1% Jan 2025 Dec 2025 Official leaderboard
28 Qwen3-VL-4B-Thinking 66.1% Oct 2025 Oct 2025 Official leaderboard
29 Qwen3-30B-A3B-Instruct Jul 2025 64.7% Jul 2025 Oct 2025 Official leaderboard
30 Qwen3-VL-30B-A3B-Instruct 62.2% Oct 2025 Oct 2025 Official leaderboard
31 Qwen2.5-VL-32B 60.1% Mar 2025 Dec 2025 Official leaderboard
32 EXAONE 4.0 32B 59.9% Jul 2025 Oct 2025 Official leaderboard
33 Llama 3.2 90B 56.9% Sep 2024 Oct 2025 Official leaderboard
34 Qwen3-VL-8B-Instruct 53.8% Oct 2025 Oct 2025 Official leaderboard
35 qwen3-4b-instruct-2507 51.5% Aug 2025 Oct 2025 Official leaderboard
36 Phi-4 51.5% Dec 2024 Dec 2025 Official leaderboard
37 Qwen3-VL-4B-Instruct 51% Oct 2025 Oct 2025 Official leaderboard
38 Qwen3-32B 47.5% Apr 2025 Oct 2025 Official leaderboard
39 Qwen3-1.7B 46.8% Apr 2025 Oct 2025 Official leaderboard
40 Qwen2.5-VL-7B 46% Jan 2025 Dec 2025 Official leaderboard
41 Qwen2.5-VL-3B 40.9% Jan 2025 Dec 2025 Official leaderboard
42 Qwen3-0.6B 32.1% Apr 2025 Oct 2025 Official leaderboard
43 Phi-4-mini-reasoning 12.6% Apr 2025 Dec 2025 Official leaderboard

Rankings by part

We read 9 parts from the benchmark’s own leaderboard in 3 groups: split, modality and domain. Each ranks models on one part of the test. Gemini 3 Pro leads every one of them. Arrows show how far a model moved from its overall rank among the models we track.

Korean specific

43 models we track have a score here.

  1. 1 Gemini 3 Pro 90.4%
  2. 2 Kimi K3 ↑8 83.6%
  3. 3 Gemma 4 31B IT 82.2%
  4. 4 GPT-5.2 ↓2 80.8%
  5. 5 DeepSeek V4 Flash 0731 ↑6 78.9%
  6. 6 DeepSeek-V4-Flash ↑6 76.1%
  7. 7 DeepSeek-V4-Pro 76.1%
  8. 8 Command A Reasoning ↑6 74.6%
  9. 9 Qwen3-VL-235B-A22B-Instruct ↑12 74%
  10. 10 Qwen3-VL-235B-A22B ↓6 72.6%

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-5.2
thinking
87.8%
Gemma 4 31B IT
non-thinking thinking
86.5%
Qwen3.5-27B
non-thinking thinking
83%
Qwen 3.6 35B-A3B
non-thinking thinking
82.9%
Qwen3.5-35B-A3B
non-thinking thinking
81.8%
Qwen3.5-9B
non-thinking thinking
74.9%

Source: Official K-MetBench 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-5.2
thinking
87.8%
Gemma 4 31B IT
non-thinking thinking
86.5%
Qwen3.5-27B
non-thinking thinking
83%
Qwen 3.6 35B-A3B
non-thinking thinking
82.9%
Qwen3.5-35B-A3B
non-thinking thinking
81.8%
Qwen3.5-9B
non-thinking thinking
74.9%

Source: Official K-MetBench leaderboard

Number one over time

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

Google DeepMind
Own leaderboard Passed 85% in Apr 2026, by Gemini 3 Pro
Gemini 3 Pro 93.7%

Source: Official K-MetBench leaderboard

Gemini 3 Pro on the official leaderboard passed 85% in November 2025

Models released before it came out already scored that high. 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 2026 Near the ceiling (85%) Gemini 3 Pro Qwen3-VL-235B-A22B Command A Reasoning Qwen2.5-VL-72B Llama 3.2 90B 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Gemini 3 Pro Qwen3-VL-235B-A22B

Source: Official K-MetBench 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 Science and reasoning

All 203 in Science and reasoning
  • GPQA Diamond In use 198 graduate-level 'Google-proof' multiple-choice questions in biology, chemistry and physics. Gemini 4 Carbon 97%
  • Terminal-Bench-Science In use 70 scientific research-workflow tasks across five science areas, run in Harbor containers. Hidden tests give the resolution rate. GPT-6 Astra 68.1%
  • BioMysteryBench In use Bioinformatics problems on anonymized real datasets that require real analysis with tools. Rubric-graded, with human-solvable and human-difficult splits. Claude Fable 5.2 83.7%
  • FrontierScience In use Models solve physics, chemistry, and biology problems in an Olympiad track graded by short answer and a Research track graded by 10-point rubrics. Intern-S2-397B 87%
  • HiPhO In use Models solve theory problems from 2024 and 2025 physics olympiad exams, some with diagrams. Official marking schemes grade answers and steps, mapped to medal levels. Gemini 3 Pro 78.3%
  • LABBench2 In use 11 biology research sub-tasks (literature, figures, tables, patents, trials and more) with tool access. Macro-averaged accuracy. GPT-5.2 60.9%

Data audit

All 9 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
  • Pass Independent runs Scores from the official leaderboard (65 entries).
  • Pass Models scored 43 models we track, plus 22 other entries on its leaderboard.
  • Pass Newest score The newest is from July 2026, for DeepSeek V4 Flash 0731.
  • Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is Kimi K3 (#25).
  • Note Headroom Saturated. The best is 93.7%, past 85% since November 2025.
Leaderboard import 3 pass
  • Pass Importer We read the board on kmetbench.github.io: 49 rows for models we track and 22 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