Benchmark Almanac

SkillsBench

GPT-5.5 67.3% leads since April 2026
Qwen 3.8 Max 70.2% leads since August 2026
In use

86-87 containerized expert tasks in 11 domains, run with and without Agent Skills. Verifier tests give pass rate and skill lift.

Paired runs with and without skills isolate the effect of skills. SkillsBench 1.1 (June 2026) pins 87 tasks in 8 domains, up from 86 in the first paper version. The task set is the same as v1.0.

Released
February 2026
Kind
Terminal tasks
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Strongest model tested
GPT-5.5 #20 on the capability index
Last in a launch post
November 2026
Last model added
November 2026
Scores refreshed
November 12th, 2026

In the authors’ words

Agent Skills are structured packages of procedural knowledge that augment large language model (LLM) agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark whose current inventory contains 87 tasks across 8 domains paired with curated Skills and deterministic verifiers. Our latest aggregate evaluation runs the 87-task benchmark under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

Abstract of SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks (arXiv:2602.12670, CC0)
22
Models scored on its own leaderboard
13
Launch posts list it, from 6 labs
2
System cards report it, from 2 labs

GPT-5.5 leads with 67.3% among models released by July 2026

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best.

Qwen 3.8 Max leads with 70.2%, claims included

Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.

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:

22 models. Scroll sideways for the rest.

  1. 67.3 GPT-5.5
  2. 61.2 Claude Opus 4.7
  3. 60.8 Gemini 3.1 Pro
  4. 58.4 GLM-5.1
  5. 55.9 Tencent Hy3
  6. 54.6 Gemini 3 Flash
  7. 54.1 Claude Opus 4.8
  8. 54.0 Kimi K2.6
  9. 53.0 MiniMax-M3
  10. 51.7 GPT-5.2
  11. 50.2 Claude Opus 4.6
  12. 50.1 DeepSeek-V4-Pro
  13. 49.0 Claude Opus 4.5
  14. 48.2 Gemini 3.5 Flash
  15. 47.2 Claude Sonnet 4.6
  16. 44.7 DeepSeek-V4-Flash
  17. 41.7 Grok 4.3
  18. 41.4 GPT-5.4 Mini
  19. 36.2 Claude Sonnet 4.5
  20. 34.9 MiniMax-M2.7
  21. 30.1 Claude Haiku 4.5
  22. 20.1 Gemini 3.1 Flash-Lite
  1. 1 GPT-5.5 67.3%
  2. 2 Claude Opus 4.7 61.2%
  3. 3 Gemini 3.1 Pro 60.8%
  4. 4 GLM-5.1 58.4%
  5. 5 Tencent Hy3 55.9%
  6. 6 Gemini 3 Flash 54.6%
  7. 7 Claude Opus 4.8 54.1%
  8. 8 Kimi K2.6 54%
  9. 9 MiniMax-M3 53%
  10. 10 GPT-5.2 51.7%
  11. 11 Claude Opus 4.6 50.2%
  12. 12 DeepSeek-V4-Pro 50.1%
  13. 13 Claude Opus 4.5 49%
  14. 14 Gemini 3.5 Flash 48.2%
  15. 15 Claude Sonnet 4.6 47.2%
  16. 16 DeepSeek-V4-Flash 44.7%
  17. 17 Grok 4.3 41.7%
  18. 18 GPT-5.4 Mini 41.4%
  19. 19 Claude Sonnet 4.5 36.2%
  20. 20 MiniMax-M2.7 34.9%
  21. 21 Claude Haiku 4.5 30.1%
  22. 22 Gemini 3.1 Flash-Lite 20.1%

Source: Official SkillsBench leaderboard

35 models. Scroll sideways for the rest.

  1. 70.2 Qwen 3.8 Max
  2. 68.7 Ling-3.1-flash
  3. 67.3 GPT-5.5
  4. 63.3 LongCat-2.5
  5. 62.9 Tencent Hy4 preview
  6. 61.2 Claude Opus 4.7
  7. 60.8 Gemini 3.1 Pro
  8. 60.4 Seed2.1 Pro
  9. 59.2 Qwen3.7-Max
  10. 58.4 GLM-5.1
  11. 55.9 Tencent Hy3
  12. 55.6 Qwen 3.6 Max (Preview)
  13. 54.9 Qwen3.7-Plus
  14. 54.6 Gemini 3 Flash
  15. 54.1 Claude Opus 4.8
  16. 54.0 Kimi K2.6
  17. 53.0 MiniMax-M3
  18. 51.7 GPT-5.2
  19. 50.2 Claude Opus 4.6
  20. 50.1 DeepSeek-V4-Pro
  21. 49.0 Claude Opus 4.5
  22. 48.2 Qwen3.6 27B
  23. 48.2 Gemini 3.5 Flash
  24. 47.2 Claude Sonnet 4.6
  25. 45.7 Qwen 3.6 Plus
  26. 44.7 DeepSeek-V4-Flash
  27. 44.3 Muse Glimmer
  28. 43.7 Seed2.0 Lite (0428)
  29. 41.7 Grok 4.3
  30. 41.4 GPT-5.4 Mini
  31. 36.2 Claude Sonnet 4.5
  32. 34.9 MiniMax-M2.7
  33. 30.1 Claude Haiku 4.5
  34. 28.7 Qwen 3.6 35B-A3B
  35. 20.1 Gemini 3.1 Flash-Lite
  1. 1 Qwen 3.8 Max 70.2%
  2. 2 Ling-3.1-flash 68.7%
  3. 3 GPT-5.5 67.3%
  4. 4 LongCat-2.5 63.3%
  5. 5 Tencent Hy4 preview 62.9%
  6. 6 Claude Opus 4.7 61.2%
  7. 7 Gemini 3.1 Pro 60.8%
  8. 8 Seed2.1 Pro 60.4%
  9. 9 Qwen3.7-Max 59.2%
  10. 10 GLM-5.1 58.4%
  11. 11 Tencent Hy3 55.9%
  12. 12 Qwen 3.6 Max (Preview) 55.6%
  13. 13 Qwen3.7-Plus 54.9%
  14. 14 Gemini 3 Flash 54.6%
  15. 15 Claude Opus 4.8 54.1%
  16. 16 Kimi K2.6 54%
  17. 17 MiniMax-M3 53%
  18. 18 GPT-5.2 51.7%
  19. 19 Claude Opus 4.6 50.2%
  20. 20 DeepSeek-V4-Pro 50.1%
  21. 21 Claude Opus 4.5 49%
  22. 22 Qwen3.6 27B 48.2%
  23. 23 Gemini 3.5 Flash 48.2%
  24. 24 Claude Sonnet 4.6 47.2%
  25. 25 Qwen 3.6 Plus 45.7%
  26. 26 DeepSeek-V4-Flash 44.7%
  27. 27 Muse Glimmer 44.3%
  28. 28 Seed2.0 Lite (0428) 43.7%
  29. 29 Grok 4.3 41.7%
  30. 30 GPT-5.4 Mini 41.4%
  31. 31 Claude Sonnet 4.5 36.2%
  32. 32 MiniMax-M2.7 34.9%
  33. 33 Claude Haiku 4.5 30.1%
  34. 34 Qwen 3.6 35B-A3B 28.7%
  35. 35 Gemini 3.1 Flash-Lite 20.1%

Sources: Labs’ model cards and official SkillsBench leaderboard

#ModelScoreSettingDateSource
1 GPT-5.5 67.3% OpenHands Apr 2026 Official leaderboard
2 Claude Opus 4.7 61.2% Claude Code Apr 2026 Official leaderboard
3 Gemini 3.1 Pro 60.8% Gemini CLI Feb 2026 Official leaderboard
4 GLM-5.1 58.4% OpenHands Apr 2026 Official leaderboard
5 Tencent Hy3 55.9% Claude Code Jul 2026 Official leaderboard
6 Gemini 3 Flash 54.6% Gemini CLI Dec 2025 Official leaderboard
7 Claude Opus 4.8 54.1% OpenHands May 2026 Official leaderboard
8 Kimi K2.6 54% OpenHands Apr 2026 Official leaderboard
9 MiniMax-M3 53% OpenHands Jun 2026 Official leaderboard
10 GPT-5.2 51.7% Codex Dec 2025 Official leaderboard
11 Claude Opus 4.6 50.2% Claude Code Feb 2026 Official leaderboard
12 DeepSeek-V4-Pro 50.1% OpenHands Apr 2026 Official leaderboard
13 Claude Opus 4.5 49% Claude Code Nov 2025 Official leaderboard
14 Gemini 3.5 Flash 48.2% OpenHands May 2026 Official leaderboard
15 Claude Sonnet 4.6 47.2% OpenHands Feb 2026 Official leaderboard
16 DeepSeek-V4-Flash 44.7% OpenHands Apr 2026 Official leaderboard
17 Grok 4.3 41.7% OpenHands Apr 2026 Official leaderboard
18 GPT-5.4 Mini 41.4% OpenHands Mar 2026 Official leaderboard
19 Claude Sonnet 4.5 36.2% Claude Code Sep 2025 Official leaderboard
20 MiniMax-M2.7 34.9% OpenHands Mar 2026 Official leaderboard
21 Claude Haiku 4.5 30.1% Claude Code Oct 2025 Official leaderboard
22 Gemini 3.1 Flash-Lite 20.1% OpenHands Mar 2026 Official leaderboard
#ModelScoreSettingDateSource
1 Qwen 3.8 Max 70.2% Aug 2026 Lab model card
2 Ling-3.1-flash 68.7% Oct 2026 Lab model card
3 GPT-5.5 67.3% OpenHands Apr 2026 Official leaderboard
4 LongCat-2.5 63.3% Nov 2026 Lab model card
5 Tencent Hy4 preview 62.9% Aug 2026 Lab model card
6 Claude Opus 4.7 61.2% Claude Code Apr 2026 Official leaderboard
7 Gemini 3.1 Pro 60.8% Gemini CLI Feb 2026 Official leaderboard
8 Seed2.1 Pro 60.4% Jun 2026 Lab model card
9 Qwen3.7-Max 59.2% May 2026 Lab model card
10 GLM-5.1 58.4% OpenHands Apr 2026 Official leaderboard
11 Tencent Hy3 55.9% Claude Code Jul 2026 Official leaderboard
12 Qwen 3.6 Max (Preview) 55.6% Apr 2026 Lab model card
13 Qwen3.7-Plus 54.9% Jun 2026 Lab model card
14 Gemini 3 Flash 54.6% Gemini CLI Dec 2025 Official leaderboard
15 Claude Opus 4.8 54.1% OpenHands May 2026 Official leaderboard
16 Kimi K2.6 54% OpenHands Apr 2026 Official leaderboard
17 MiniMax-M3 53% OpenHands Jun 2026 Official leaderboard
18 GPT-5.2 51.7% Codex Dec 2025 Official leaderboard
19 Claude Opus 4.6 50.2% Claude Code Feb 2026 Official leaderboard
20 DeepSeek-V4-Pro 50.1% OpenHands Apr 2026 Official leaderboard
21 Claude Opus 4.5 49% Claude Code Nov 2025 Official leaderboard
22 Qwen3.6 27B 48.2% Apr 2026 Lab model card
23 Gemini 3.5 Flash 48.2% OpenHands May 2026 Official leaderboard
24 Claude Sonnet 4.6 47.2% OpenHands Feb 2026 Official leaderboard
25 Qwen 3.6 Plus 45.7% Apr 2026 Lab model card
26 DeepSeek-V4-Flash 44.7% OpenHands Apr 2026 Official leaderboard
27 Muse Glimmer 44.3% Aug 2026 Lab model card
28 Seed2.0 Lite (0428) 43.7% Apr 2026 Lab model card
29 Grok 4.3 41.7% OpenHands Apr 2026 Official leaderboard
30 GPT-5.4 Mini 41.4% OpenHands Mar 2026 Official leaderboard
31 Claude Sonnet 4.5 36.2% Claude Code Sep 2025 Official leaderboard
32 MiniMax-M2.7 34.9% OpenHands Mar 2026 Official leaderboard
33 Claude Haiku 4.5 30.1% Claude Code Oct 2025 Official leaderboard
34 Qwen 3.6 35B-A3B 28.7% Apr 2026 Lab model card
35 Gemini 3.1 Flash-Lite 20.1% OpenHands Mar 2026 Official leaderboard

Labs report scores for 14 models in their cards. Switch the scores above to rank them together.

4 models have held number one on the official leaderboard since its release in February 2026 5 models have held number one with claims included since its release in February 2026

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.

Google DeepMind Anthropic OpenAI Qwen ByteDance Seed
Own leaderboard Not tested since Jul 2026
GPT-5.5 67.3%
Lab claims 5 leaders
Qwen 3.8 Max 70.2%

Sources: Official SkillsBench leaderboard and labs’ model cards

Scores over time

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

On the benchmark’s own leaderboard Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Qwen 3.8 Max GPT-5.5 Claude Opus 4.7 Gemini 3 Flash Claude Opus 4.5 Claude Sonnet 4.5 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Qwen 3.8 Max Claude Opus 4.7

Sources: Official SkillsBench leaderboard and labs’ model cards

5 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

Qwen
ByteDance Seed
Tencent
Meta
Ant Group
Meituan
DateLaunch postScoreAs the post writes it
Nov 2, 2026 LongCat-2.5 63.3 SkillsBench
Oct 13, 2026 Ling-3.1-flash 68.7 SkillsBench
Aug 28, 2026 Tencent Hy4 preview 62.9 SkillsBench (79, text-only)
Aug 10, 2026 Muse Glimmer 44.3 SkillsBench (with skills)
Aug 3, 2026 Qwen 3.8 Max 70.2 SkillsBench
Jul 2, 2026 Tencent Hy3 55.3 SkillsBench (79, text-only)
Jun 1, 2026 Qwen3.7-Plus 54.9 Skillsbench (OpenCode, 78 tasks, avg of 5)
May 20, 2026 Qwen3.7-Max 59.2 Skillsbench
Apr 28, 2026 Seed2.0 Lite (0428) 43.7 SkillsBench
Apr 22, 2026 Qwen3.6 27B 48.2 SkillsBench Avg5 (OpenCode, 78 tasks)
Apr 20, 2026 Qwen 3.6 Max (Preview) 55.6 SkillsBench Agent Skills
Apr 15, 2026 Qwen 3.6 35B-A3B 28.7 SkillsBench Avg5
Apr 2, 2026 Qwen 3.6 Plus 45.7 SkillsBench Avg5

2 system cards report it, from 2 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
Oct 19, 2026 Muse Glimmer technical report 44.3 SkillsBench, high (Tables 2 and 5) Capabilities
Jun 23, 2026 Seed2.1 Pro model card 60.4 SkillsBench, stated in the 2.3.2 text Capabilities

One lab claim next to the leaderboard

Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. Every benchmark.

Model Lab says Board says Gap
Tencent Hy3 55.3% SkillsBench (79, text-only) 55.9% Leaderboard, Claude Code −0.6 pts

More in Terminal agents

All 19 in Terminal agents
  • Terminal-Bench Hard In use Agents complete 44 hard tasks from Terminal-Bench Core in the Terminus 2 harness. Artificial Analysis scores pass@1 averaged over 3 repeats. Gemma 4 31B IT 36%
  • Harbor-Index In use Agents solve about 80 hard tasks drawn from 29 benchmarks run through Harbor adapters, such as HLE, GSO and SWE-bench Pro. Each task keeps its source verifier. GPT-5.5 28.1%
  • Terminal-Bench 2.1 Extended Not in model cards Agents complete long command-line tasks in containers, scored by verification tests. Mercor adds 99 new expert-written tasks to the 89 public ones and reports both scores. Muse Wonder 54%
  • Terminal-Bench 4.1 Not in model cards The Terminal-Bench 4.0 tasks with hardened verifiers: agents complete 66 hard tasks in containerized terminals across software, ML, science, security, and operations. Test scripts verify each result, scored as resolution rate. Claude Opus 5.5 65.8%
  • o11y-bench Not in model cards Agents query metrics, logs and traces in a live Grafana stack, investigate incidents and fix dashboards. Grading checks claims against facts read from the same environment. Claude Opus 5.5 93.8%
  • CompileBench Not in model cards Agents build real open-source projects through dependency problems, legacy toolchains and complex build systems, for example compiling 22-year-old code. Claude Opus 4.1 100%