Benchmark Almanac

C3-Bench

Gemini-2.0-Thinking (snapshot unspecified) 14.4% tops its leaderboard Claude Sonnet 4 12.5% leads the models we track
Hunyuan-A13B-Instruct 63.5% leads since June 2025
Dropped

Agents complete multi-task tool use dialogues built to stress tool dependencies, long context dependencies and frequent switches between policy types.

The authors moved C3-Bench to WildToolBench. The maintained board evaluates 57 configurations on 256 sessions. Session accuracy is the headline score.

Made by
Tencent Hunyuan
Released
May 2025
Kind
Function calling
Level
University
Scored
Checked automatically
Answers
Agent task
Tasks
All public
Strongest model tested
GPT-5 #70 on the capability index
Last in a launch post
June 2025, dropped before it saturated
Last model added
August 2025
Scores refreshed
November 12th, 2026
50
Entries on its own leaderboard, 30 models we track
1
Launch post lists it, from 1 lab
1
System card reports it, from 1 lab

Gemini-2.0-Thinking (snapshot unspecified) tops the leaderboard with 14.4%

Best score per entry from the benchmark's own leaderboard, read on November 12th, 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. The best model we track is Claude Sonnet 4, at 12.5%.

Hunyuan-A13B-Instruct leads with 63.5%, 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 August 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

50 entries. Scroll sideways for the rest.

  1. 14.4 Gemini-2.0-Thinking (snapshot unspecified)
  2. 14.1 Gemini-2.5-Pro (snapshot unspecified)
  3. 12.5 Claude Sonnet 4
  4. 12.1 o1
  5. 12.1 GLM-4.5
  6. 11.7 GPT-4o (snapshot unspecified)
  7. 11.3 Claude 3.7 Sonnet
  8. 10.6 Kimi-K2 (snapshot unspecified)
  9. 10.2 o3
  10. 10.2 Grok 4
  11. 9.77 Qwen3-30B-A3B
  12. 9.38 DeepSeek-R1 (snapshot unspecified)
  13. 9.38 DeepSeek-V3 (snapshot unspecified)
  14. 9.38 Qwen3-14B
  15. 9.38 Claude Opus 4.1
  16. 8.98 GPT-4.1
  17. 8.98 Qwen3-8B
  18. 8.98 DeepSeek-V3.1
  19. 8.59 Qwen3-32B
  20. 8.20 xLAM-2-32b-fc-r
  21. 7.81 Qwen2.5-14B
  22. 7.81 Llama-xLAM-2-70b-fc-r
  23. 7.81 Qwen3-4B
  24. 7.42 ToolACE2-8B
  25. 7.03 Doubao-1.6 (snapshot unspecified)
  26. 7.03 Mistral-Large (snapshot unspecified)
  27. 6.64 Doubao-1.5-Thinking (snapshot unspecified)
  28. 6.25 Qwen2.5-72B
  29. 6.25 Qwen3-1.7B
  30. 5.86 Qwen2.5-32B
  31. 5.86 GPT-5
  32. 5.08 Llama-xLAM-2-8b-fc-r
  33. 4.69 Hammer2.1-7B
  34. 4.69 Watt-8B
  35. 4.30 Qwen2.5-3B
  36. 4.30 Qwen2.5-7B
  37. 3.91 Qwen2.5-1.5B
  38. 3.91 Qwen3-0.6B
  39. 3.13 Doubao-1.6-Thinking (snapshot unspecified)
  40. 2.34 xLAM-2-1b-fc-r
  41. 1.17 xLAM-2-3b-fc-r
  42. 0.78 Doubao-1.5 (snapshot unspecified)
  43. 0.78 Qwen2.5-0.5B
  44. 0.39 Hammer2.1-3B
  45. 0.39 Llama-3.3-1B-Instruct
  46. 0.39 Llama-3.3-3B-Instruct
  47. 0.0 Hammer2.1-0.5B
  48. 0.0 Hammer2.1-1.5B
  49. 0.0 Llama-3.3-8B-Instruct
  50. 0.0 Llama 3.3 70B Instruct
  1. 1 Gemini-2.0-Thinking (snapshot unspecified) 14.4%
  2. 2 Gemini-2.5-Pro (snapshot unspecified) 14.1%
  3. 3 Claude Sonnet 4 12.5%
  4. 4 o1 12.1%
  5. 5 GLM-4.5 12.1%
  6. 6 GPT-4o (snapshot unspecified) 11.7%
  7. 7 Claude 3.7 Sonnet 11.3%
  8. 8 Kimi-K2 (snapshot unspecified) 10.6%
  9. 9 o3 10.2%
  10. 10 Grok 4 10.2%
  11. 11 Qwen3-30B-A3B non-thinking 9.77%
  12. 12 DeepSeek-R1 (snapshot unspecified) 9.38%
  13. 13 DeepSeek-V3 (snapshot unspecified) 9.38%
  14. 14 Qwen3-14B thinking 9.38%
  15. 15 Claude Opus 4.1 9.38%
  16. 16 GPT-4.1 8.98%
  17. 17 Qwen3-8B thinking 8.98%
  18. 18 DeepSeek-V3.1 8.98%
  19. 19 Qwen3-32B non-thinking 8.59%
  20. 20 xLAM-2-32b-fc-r 8.20%
  21. 21 Qwen2.5-14B 7.81%
  22. 22 Llama-xLAM-2-70b-fc-r 7.81%
  23. 23 Qwen3-4B thinking 7.81%
  24. 24 ToolACE2-8B 7.42%
  25. 25 Doubao-1.6 (snapshot unspecified) 7.03%
  26. 26 Mistral-Large (snapshot unspecified) 7.03%
  27. 27 Doubao-1.5-Thinking (snapshot unspecified) 6.64%
  28. 28 Qwen2.5-72B 6.25%
  29. 29 Qwen3-1.7B thinking 6.25%
  30. 30 Qwen2.5-32B 5.86%
  31. 31 GPT-5 5.86%
  32. 32 Llama-xLAM-2-8b-fc-r 5.08%
  33. 33 Hammer2.1-7B 4.69%
  34. 34 Watt-8B 4.69%
  35. 35 Qwen2.5-3B 4.30%
  36. 36 Qwen2.5-7B 4.30%
  37. 37 Qwen2.5-1.5B 3.91%
  38. 38 Qwen3-0.6B thinking 3.91%
  39. 39 Doubao-1.6-Thinking (snapshot unspecified) 3.13%
  40. 40 xLAM-2-1b-fc-r 2.34%
  41. 41 xLAM-2-3b-fc-r 1.17%
  42. 42 Doubao-1.5 (snapshot unspecified) 0.78%
  43. 43 Qwen2.5-0.5B 0.78%
  44. 44 Hammer2.1-3B 0.39%
  45. 45 Llama-3.3-1B-Instruct 0.39%
  46. 46 Llama-3.3-3B-Instruct 0.39%
  47. 47 Hammer2.1-0.5B 0%
  48. 48 Hammer2.1-1.5B 0%
  49. 49 Llama-3.3-8B-Instruct 0%
  50. 50 Llama 3.3 70B Instruct 0%

Source: Official C3-Bench leaderboard

31 models. Scroll sideways for the rest.

  1. 63.5 Hunyuan-A13B-Instruct
  2. 12.5 Claude Sonnet 4
  3. 12.1 o1
  4. 12.1 GLM-4.5
  5. 11.3 Claude 3.7 Sonnet
  6. 10.2 o3
  7. 10.2 Grok 4
  8. 9.77 Qwen3-30B-A3B
  9. 9.38 Qwen3-14B
  10. 9.38 Claude Opus 4.1
  11. 8.98 GPT-4.1
  12. 8.98 Qwen3-8B
  13. 8.98 DeepSeek-V3.1
  14. 8.59 Qwen3-32B
  15. 8.20 xLAM-2-32b-fc-r
  16. 7.81 Qwen2.5-14B
  17. 7.81 Llama-xLAM-2-70b-fc-r
  18. 7.81 Qwen3-4B
  19. 6.25 Qwen2.5-72B
  20. 6.25 Qwen3-1.7B
  21. 5.86 Qwen2.5-32B
  22. 5.86 GPT-5
  23. 5.08 Llama-xLAM-2-8b-fc-r
  24. 4.30 Qwen2.5-3B
  25. 4.30 Qwen2.5-7B
  26. 3.91 Qwen2.5-1.5B
  27. 3.91 Qwen3-0.6B
  28. 2.34 xLAM-2-1b-fc-r
  29. 1.17 xLAM-2-3b-fc-r
  30. 0.78 Qwen2.5-0.5B
  31. 0.0 Llama 3.3 70B Instruct
  1. 1 Hunyuan-A13B-Instruct 63.5%
  2. 2 Claude Sonnet 4 12.5%
  3. 3 o1 12.1%
  4. 4 GLM-4.5 12.1%
  5. 5 Claude 3.7 Sonnet 11.3%
  6. 6 o3 10.2%
  7. 7 Grok 4 10.2%
  8. 8 Qwen3-30B-A3B non-thinking 9.77%
  9. 9 Qwen3-14B thinking 9.38%
  10. 10 Claude Opus 4.1 9.38%
  11. 11 GPT-4.1 8.98%
  12. 12 Qwen3-8B thinking 8.98%
  13. 13 DeepSeek-V3.1 8.98%
  14. 14 Qwen3-32B non-thinking 8.59%
  15. 15 xLAM-2-32b-fc-r 8.20%
  16. 16 Qwen2.5-14B 7.81%
  17. 17 Llama-xLAM-2-70b-fc-r 7.81%
  18. 18 Qwen3-4B thinking 7.81%
  19. 19 Qwen2.5-72B 6.25%
  20. 20 Qwen3-1.7B thinking 6.25%
  21. 21 Qwen2.5-32B 5.86%
  22. 22 GPT-5 5.86%
  23. 23 Llama-xLAM-2-8b-fc-r 5.08%
  24. 24 Qwen2.5-3B 4.30%
  25. 25 Qwen2.5-7B 4.30%
  26. 26 Qwen2.5-1.5B 3.91%
  27. 27 Qwen3-0.6B thinking 3.91%
  28. 28 xLAM-2-1b-fc-r 2.34%
  29. 29 xLAM-2-3b-fc-r 1.17%
  30. 30 Qwen2.5-0.5B 0.78%
  31. 31 Llama 3.3 70B Instruct 0%

Sources: Labs’ model cards and official C3-Bench leaderboard

#EntryScoreSettingDateSource
1 Gemini-2.0-Thinking (snapshot unspecified) 14.4% Official leaderboard
2 Gemini-2.5-Pro (snapshot unspecified) 14.1% Official leaderboard
3 Claude Sonnet 4 12.5% May 2025 Official leaderboard
4 o1 12.1% Dec 2024 Official leaderboard
5 GLM-4.5 12.1% Aug 2025 Official leaderboard
6 GPT-4o (snapshot unspecified) 11.7% Official leaderboard
7 Claude 3.7 Sonnet 11.3% Feb 2025 Official leaderboard
8 Kimi-K2 (snapshot unspecified) 10.6% Official leaderboard
9 o3 10.2% Apr 2025 Official leaderboard
10 Grok 4 10.2% Jul 2025 Official leaderboard
11 Qwen3-30B-A3B 9.77% non-thinking Apr 2025 Official leaderboard
12 DeepSeek-R1 (snapshot unspecified) 9.38% Official leaderboard
13 DeepSeek-V3 (snapshot unspecified) 9.38% Official leaderboard
14 Qwen3-14B 9.38% thinking Apr 2025 Official leaderboard
15 Claude Opus 4.1 9.38% Aug 2025 Official leaderboard
16 GPT-4.1 8.98% Apr 2025 Official leaderboard
17 Qwen3-8B 8.98% thinking Apr 2025 Official leaderboard
18 DeepSeek-V3.1 8.98% Aug 2025 Official leaderboard
19 Qwen3-32B 8.59% non-thinking Apr 2025 Official leaderboard
20 xLAM-2-32b-fc-r 8.20% Mar 2025 Official leaderboard
21 Qwen2.5-14B 7.81% Sep 2024 Official leaderboard
22 Llama-xLAM-2-70b-fc-r 7.81% Mar 2025 Official leaderboard
23 Qwen3-4B 7.81% thinking Apr 2025 Official leaderboard
24 ToolACE2-8B 7.42% Official leaderboard
25 Doubao-1.6 (snapshot unspecified) 7.03% Official leaderboard
26 Mistral-Large (snapshot unspecified) 7.03% Official leaderboard
27 Doubao-1.5-Thinking (snapshot unspecified) 6.64% Official leaderboard
28 Qwen2.5-72B 6.25% Sep 2024 Official leaderboard
29 Qwen3-1.7B 6.25% thinking Apr 2025 Official leaderboard
30 Qwen2.5-32B 5.86% Sep 2024 Official leaderboard
31 GPT-5 5.86% Aug 2025 Official leaderboard
32 Llama-xLAM-2-8b-fc-r 5.08% Mar 2025 Official leaderboard
33 Hammer2.1-7B 4.69% Official leaderboard
34 Watt-8B 4.69% Official leaderboard
35 Qwen2.5-3B 4.30% Sep 2024 Official leaderboard
36 Qwen2.5-7B 4.30% Sep 2024 Official leaderboard
37 Qwen2.5-1.5B 3.91% Sep 2024 Official leaderboard
38 Qwen3-0.6B 3.91% thinking Apr 2025 Official leaderboard
39 Doubao-1.6-Thinking (snapshot unspecified) 3.13% Official leaderboard
40 xLAM-2-1b-fc-r 2.34% Mar 2025 Official leaderboard
41 xLAM-2-3b-fc-r 1.17% Mar 2025 Official leaderboard
42 Doubao-1.5 (snapshot unspecified) 0.78% Official leaderboard
43 Qwen2.5-0.5B 0.78% Sep 2024 Official leaderboard
44 Hammer2.1-3B 0.39% Official leaderboard
45 Llama-3.3-1B-Instruct 0.39% Official leaderboard
46 Llama-3.3-3B-Instruct 0.39% Official leaderboard
47 Hammer2.1-0.5B 0% Official leaderboard
48 Hammer2.1-1.5B 0% Official leaderboard
49 Llama-3.3-8B-Instruct 0% Official leaderboard
50 Llama 3.3 70B Instruct 0% Dec 2024 Official leaderboard
#ModelScoreSettingDateSource
1 Hunyuan-A13B-Instruct 63.5% Jun 2025 Lab model card
2 Claude Sonnet 4 12.5% May 2025 Official leaderboard
3 o1 12.1% Dec 2024 Official leaderboard
4 GLM-4.5 12.1% Aug 2025 Official leaderboard
5 Claude 3.7 Sonnet 11.3% Feb 2025 Official leaderboard
6 o3 10.2% Apr 2025 Official leaderboard
7 Grok 4 10.2% Jul 2025 Official leaderboard
8 Qwen3-30B-A3B 9.77% non-thinking Apr 2025 Official leaderboard
9 Qwen3-14B 9.38% thinking Apr 2025 Official leaderboard
10 Claude Opus 4.1 9.38% Aug 2025 Official leaderboard
11 GPT-4.1 8.98% Apr 2025 Official leaderboard
12 Qwen3-8B 8.98% thinking Apr 2025 Official leaderboard
13 DeepSeek-V3.1 8.98% Aug 2025 Official leaderboard
14 Qwen3-32B 8.59% non-thinking Apr 2025 Official leaderboard
15 xLAM-2-32b-fc-r 8.20% Mar 2025 Official leaderboard
16 Qwen2.5-14B 7.81% Sep 2024 Official leaderboard
17 Llama-xLAM-2-70b-fc-r 7.81% Mar 2025 Official leaderboard
18 Qwen3-4B 7.81% thinking Apr 2025 Official leaderboard
19 Qwen2.5-72B 6.25% Sep 2024 Official leaderboard
20 Qwen3-1.7B 6.25% thinking Apr 2025 Official leaderboard
21 Qwen2.5-32B 5.86% Sep 2024 Official leaderboard
22 GPT-5 5.86% Aug 2025 Official leaderboard
23 Llama-xLAM-2-8b-fc-r 5.08% Mar 2025 Official leaderboard
24 Qwen2.5-3B 4.30% Sep 2024 Official leaderboard
25 Qwen2.5-7B 4.30% Sep 2024 Official leaderboard
26 Qwen2.5-1.5B 3.91% Sep 2024 Official leaderboard
27 Qwen3-0.6B 3.91% thinking Apr 2025 Official leaderboard
28 xLAM-2-1b-fc-r 2.34% Mar 2025 Official leaderboard
29 xLAM-2-3b-fc-r 1.17% Mar 2025 Official leaderboard
30 Qwen2.5-0.5B 0.78% Sep 2024 Official leaderboard
31 Llama 3.3 70B Instruct 0% Dec 2024 Official leaderboard

Labs report scores for 1 model in their cards. Switch the scores above to rank them together.

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
Qwen3-30B-A3B
thinking non-thinking
9.77%
Qwen3-14B
non-thinking thinking
9.38%
Qwen3-8B
non-thinking thinking
8.98%
Qwen3-32B
thinking non-thinking
8.59%
Qwen3-4B
non-thinking thinking
7.81%
Qwen3-1.7B
non-thinking thinking
6.25%
Qwen3-0.6B
non-thinking
3.91%

Source: Official C3-Bench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Claude Sonnet 4 12.5%
o1 12.1%
GLM-4.5 12.1%
Claude 3.7 Sonnet 11.3%
o3 10.2%
Grok 4 10.2%
Qwen3-30B-A3B non-thinking 9.77%
Qwen3-14B thinking 9.38%
Claude Opus 4.1 9.38%
GPT-4.1 8.98%
Qwen3-8B thinking 8.98%
DeepSeek-V3.1 8.98%
Qwen3-32B non-thinking 8.59%
xLAM-2-32b-fc-r 8.20%
Qwen2.5-14B 7.81%
Llama-xLAM-2-70b-fc-r 7.81%
Qwen3-14B non-thinking 7.81%
Qwen3-32B thinking 7.81%
Qwen3-4B thinking 7.81%
Qwen2.5-72B 6.25%
Qwen3-30B-A3B thinking 6.25%
Qwen3-8B non-thinking 6.25%
Qwen3-1.7B thinking 6.25%
Qwen2.5-32B 5.86%
Qwen3-4B non-thinking 5.86%
GPT-5 5.86%
Llama-xLAM-2-8b-fc-r 5.08%
Qwen2.5-3B 4.30%
Qwen2.5-7B 4.30%
Qwen3-1.7B non-thinking 4.30%
Qwen2.5-1.5B 3.91%
Qwen3-0.6B thinking 3.91%
Qwen3-0.6B non-thinking 3.52%
xLAM-2-1b-fc-r 2.34%
xLAM-2-3b-fc-r 1.17%
Qwen2.5-0.5B 0.78%
Llama 3.3 70B Instruct 0%

2 models have held number one on the official leaderboard since its release in May 2025 3 models have held number one with claims included since its release in May 2025

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

OpenAI Anthropic
Own leaderboard Not tested since Aug 2025
Claude Sonnet 4 12.5%

Source: Official C3-Bench leaderboard

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% 2025 Near the ceiling (85%) Hunyuan-A13B-Instruct Claude Sonnet 4 o1 Qwen2.5-14B 0% 20% 40% 60% 80% 100% 2025 Near the ceiling (85%) Hunyuan-A13B-Instruct Claude Sonnet 4

Sources: Official C3-Bench leaderboard and labs’ model cards

No launch post has listed it since June 2025

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.

Tencent
DateLaunch postScoreAs the post writes it
Jun 27, 2025 Hunyuan-A13B-Instruct 63.5 C3-Bench

One system card reports it

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
Jun 26, 2025 Hunyuan-A13B-Instruct technical report 63.5 C3-Bench, slow-thinking (Table 3). Fast-thinking 65.4 (Table 4) Capabilities

More in Tool use

All 144 in Tool use
  • AutomationBench In use Agent completes end-to-end business workflows inside a simulated company exposing dozens of REST APIs. Graded on resulting state. Gemini 4 Carbon 53.9%
  • MCP Atlas In use Multi-step workflows over real MCP servers. Scores tool-use competency on realistic tasks. Gemini 4 Argon 90.7%
  • BFCL V4 In use Models call functions and tools, including multi-turn agentic tasks with web search and memory. Checked against expected calls and outcomes. Qwen3.7-Max 75%
  • τ³-bench Banking In use Agents act as banking customer-service reps over many turns with a simulated user, following policy. Pass^1 on the final database state. Qwen 3.8 Max 55.1%
  • Toolathlon-Verified In use 108 long-horizon tasks across 32 apps through about 600 MCP tools. Execution-based checkers inspect final environment state. GLM-5.3-Flash 78.4%
  • DeepSearchQA In use Questions that need multi-step search and synthesis. Final answers graded against references. Kimi K3 95%