Benchmark Almanac

Tax Agent Bench

Gemini 4 Hydrogen 52.3% leads since November 2026
Inkling 2 Small 65.8% leads since November 2026
In use

Agents research 391 expert-written US corporate tax questions with tools and submit one cited written answer. A judge grades it against a rubric and checks that citations are real.

Splits are 5 public, 193 licensable validation and 193 private test questions. Main metric is all-pass accuracy, with weighted pass as partial credit. 80 models on the board on October 29th, 2026.

Made by
Vals AI
Released
September 2026
Kind
Finance
Level
Expert
Scored
Graded by a model
Answers
Agent task
Tasks
Public, plus a private test set
Strongest model tested
Claude Fable 5.2 #1 on the capability index
Last in a launch post
November 2026
Last model added
November 2026
Scores refreshed
November 13th, 2026
84
Models scored on its own leaderboard
1
Launch post lists it, from 1 lab

Gemini 4 Hydrogen leads with 52.3%

Best score per model from the benchmark's own leaderboard, read on November 13th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

Inkling 2 Small leads with 65.8%, 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.

84 models. Scroll sideways for the rest.

  1. 52.3 Gemini 4 Hydrogen
  2. 51.8 Gemini 4 Carbon
  3. 50.8 Claude Fable 5.1
  4. 49.7 Qwen3.8 Max (0902)
  5. 49.2 Gemini 4 Argon
  6. 47.1 Claude Opus 5
  7. 44.5 Claude Opus 5.5
  8. 44.4 DeepSeek V4.1 Flash 1017
  9. 43.3 GLM-5.3
  10. 43.0 Muse Wonder Full
  11. 42.4 Muse Spark 1.3
  12. 42.0 Claude Sonnet 5.5
  13. 38.9 Inkling 2 Small
  14. 38.3 Claude Fable 5
  15. 38.2 Muse Ember
  16. 36.3 Muse Wonder
  17. 35.8 Grok 4.7
  18. 34.4 Tencent Hy4
  19. 33.7 Grok 4.6
  20. 33.4 Kimi K3
  21. 32.8 Muse Spark 1.2
  22. 32.6 Tencent Hy4 preview
  23. 32.4 MiMo-V2.6 Pro
  24. 32.1 MiniMax-M3.1-Flash
  25. 32.1 Gemini 3.8 Flash
  26. 32.1 Qwen 3.8 Max
  27. 32.1 Claude Fable 5.2
  28. 31.9 Claude Haiku 5.5
  29. 31.6 GLM-5.3-Flash
  30. 31.1 GPT-5.6 Sol
  31. 31.1 Ember-1
  32. 30.9 Qwen 3.8 27B
  33. 30.7 Mistral Large 4
  34. 30.4 Mistral Large 4 Preview
  35. 30.0 GPT-5.6 Terra
  36. 30.0 Claude Opus 4.8
  37. 29.4 Claude Sonnet 5
  38. 28.5 MiMo-V2.6 Flash
  39. 28.1 DeepSeek V4 Flash 0731
  40. 27.3 Step 5 Preview
  41. 26.8 DeepSeek V4 Pro 0813
  42. 26.6 DeepSeek V4.1 Flash
  43. 26.3 Muse Spark Open
  44. 25.8 Claude Sonnet 4.6
  45. 25.2 DeepSeek-V4-Pro
  46. 24.9 Kimi K3.1 Pro
  47. 24.9 Grok 4.5
  48. 24.8 GPT-5.6 Luna
  49. 23.7 Claude Opus 4.7
  50. 22.8 Gemini 3.7 Flash
  51. 21.8 GPT-5.5
  52. 21.8 GPT-6.1 Sol
  53. 21.6 Gemini 3.5 Flash
  54. 20.7 GPT-6 Astra
  55. 20.6 Ling-3.0-flash-Fin
  56. 20.6 MiniMax-M3
  57. 19.9 GPT-6 Luna
  58. 18.2 Mercury 3
  59. 18.0 Gemini 3.6 Flash
  60. 17.1 GPT-6.1 Terra
  61. 15.8 Qwen3.7-Max
  62. 15.4 Inkling-Small
  63. 15.3 Inkling
  64. 15.0 GPT-6 Sol
  65. 12.4 Kimi K2.6
  66. 11.8 Gemini 3 Flash
  67. 11.6 Nemotron 3.5 Super VL
  68. 11.5 MiMo-V2.5-Pro
  69. 10.5 Qwen3.7-Plus
  70. 9.29 Gemini 3.1 Pro
  71. 8.81 GPT-5.4 Mini
  72. 8.81 Beam 1.0
  73. 7.77 MiMo-V2.5
  74. 7.68 Qwen 3.6 Plus
  75. 7.25 Gemini 3.5 Flash-Lite
  76. 6.95 Grok 4.20
  77. 6.87 MiniMax-M2.7
  78. 6.18 Grok 4.3
  79. 5.86 Claude Haiku 4.5
  80. 4.97 GPT-5.4 Nano
  81. 4.48 Mistral Medium 3.5
  82. 3.11 Gemini 3.1 Flash-Lite
  83. 2.59 Nemotron 3.5 Lightning 30B-A3B
  84. 2.49 Mercury 2.5
  1. 1 Gemini 4 Hydrogen 52.3%
  2. 2 Gemini 4 Carbon 51.8%
  3. 3 Claude Fable 5.1 50.8%
  4. 4 Qwen3.8 Max (0902) 49.7%
  5. 5 Gemini 4 Argon 49.2%
  6. 6 Claude Opus 5 47.1%
  7. 7 Claude Opus 5.5 44.5%
  8. 8 DeepSeek V4.1 Flash 1017 44.4%
  9. 9 GLM-5.3 43.3%
  10. 10 Muse Wonder Full 43%
  11. 11 Muse Spark 1.3 max 42.4%
  12. 12 Claude Sonnet 5.5 42%
  13. 13 Inkling 2 Small 38.9%
  14. 14 Claude Fable 5 38.3%
  15. 15 Muse Ember 38.2%
  16. 16 Muse Wonder 36.3%
  17. 17 Grok 4.7 35.8%
  18. 18 Tencent Hy4 34.4%
  19. 19 Grok 4.6 33.7%
  20. 20 Kimi K3 33.4%
  21. 21 Muse Spark 1.2 32.8%
  22. 22 Tencent Hy4 preview 32.6%
  23. 23 MiMo-V2.6 Pro 32.4%
  24. 24 MiniMax-M3.1-Flash 32.1%
  25. 25 Gemini 3.8 Flash 32.1%
  26. 26 Qwen 3.8 Max 32.1%
  27. 27 Claude Fable 5.2 32.1%
  28. 28 Claude Haiku 5.5 31.9%
  29. 29 GLM-5.3-Flash 31.6%
  30. 30 GPT-5.6 Sol 31.1%
  31. 31 Ember-1 31.1%
  32. 32 Qwen 3.8 27B 30.9%
  33. 33 Mistral Large 4 30.7%
  34. 34 Mistral Large 4 Preview 30.4%
  35. 35 GPT-5.6 Terra 30%
  36. 36 Claude Opus 4.8 30%
  37. 37 Claude Sonnet 5 29.4%
  38. 38 MiMo-V2.6 Flash 28.5%
  39. 39 DeepSeek V4 Flash 0731 28.1%
  40. 40 Step 5 Preview 27.3%
  41. 41 DeepSeek V4 Pro 0813 26.8%
  42. 42 DeepSeek V4.1 Flash 26.6%
  43. 43 Muse Spark Open 26.3%
  44. 44 Claude Sonnet 4.6 25.8%
  45. 45 DeepSeek-V4-Pro 25.2%
  46. 46 Kimi K3.1 Pro 24.9%
  47. 47 Grok 4.5 24.9%
  48. 48 GPT-5.6 Luna 24.8%
  49. 49 Claude Opus 4.7 23.7%
  50. 50 Gemini 3.7 Flash 22.8%
  51. 51 GPT-5.5 21.8%
  52. 52 GPT-6.1 Sol 21.8%
  53. 53 Gemini 3.5 Flash 21.6%
  54. 54 GPT-6 Astra 20.7%
  55. 55 Ling-3.0-flash-Fin 20.6%
  56. 56 MiniMax-M3 20.6%
  57. 57 GPT-6 Luna 19.9%
  58. 58 Mercury 3 18.2%
  59. 59 Gemini 3.6 Flash 18%
  60. 60 GPT-6.1 Terra 17.1%
  61. 61 Qwen3.7-Max 15.8%
  62. 62 Inkling-Small 15.4%
  63. 63 Inkling 15.3%
  64. 64 GPT-6 Sol 15%
  65. 65 Kimi K2.6 12.4%
  66. 66 Gemini 3 Flash 11.8%
  67. 67 Nemotron 3.5 Super VL 11.6%
  68. 68 MiMo-V2.5-Pro 11.5%
  69. 69 Qwen3.7-Plus 10.5%
  70. 70 Gemini 3.1 Pro 9.29%
  71. 71 GPT-5.4 Mini 8.81%
  72. 72 Beam 1.0 8.81%
  73. 73 MiMo-V2.5 7.77%
  74. 74 Qwen 3.6 Plus 7.68%
  75. 75 Gemini 3.5 Flash-Lite 7.25%
  76. 76 Grok 4.20 reasoning 6.95%
  77. 77 MiniMax-M2.7 6.87%
  78. 78 Grok 4.3 6.18%
  79. 79 Claude Haiku 4.5 thinking 5.86%
  80. 80 GPT-5.4 Nano 4.97%
  81. 81 Mistral Medium 3.5 4.48%
  82. 82 Gemini 3.1 Flash-Lite 3.11%
  83. 83 Nemotron 3.5 Lightning 30B-A3B 2.59%
  84. 84 Mercury 2.5 2.49%

Source: Official Tax Agent Bench leaderboard

84 models. Scroll sideways for the rest.

  1. 65.8 Inkling 2 Small
  2. 52.3 Gemini 4 Hydrogen
  3. 51.8 Gemini 4 Carbon
  4. 50.8 Claude Fable 5.1
  5. 49.7 Qwen3.8 Max (0902)
  6. 49.2 Gemini 4 Argon
  7. 47.1 Claude Opus 5
  8. 44.5 Claude Opus 5.5
  9. 44.4 DeepSeek V4.1 Flash 1017
  10. 43.3 GLM-5.3
  11. 43.0 Muse Wonder Full
  12. 42.4 Muse Spark 1.3
  13. 42.0 Claude Sonnet 5.5
  14. 38.3 Claude Fable 5
  15. 38.2 Muse Ember
  16. 36.3 Muse Wonder
  17. 35.8 Grok 4.7
  18. 34.4 Tencent Hy4
  19. 33.7 Grok 4.6
  20. 33.4 Kimi K3
  21. 32.8 Muse Spark 1.2
  22. 32.6 Tencent Hy4 preview
  23. 32.4 MiMo-V2.6 Pro
  24. 32.1 MiniMax-M3.1-Flash
  25. 32.1 Gemini 3.8 Flash
  26. 32.1 Qwen 3.8 Max
  27. 32.1 Claude Fable 5.2
  28. 31.9 Claude Haiku 5.5
  29. 31.6 GLM-5.3-Flash
  30. 31.1 GPT-5.6 Sol
  31. 31.1 Ember-1
  32. 30.9 Qwen 3.8 27B
  33. 30.7 Mistral Large 4
  34. 30.4 Mistral Large 4 Preview
  35. 30.0 GPT-5.6 Terra
  36. 30.0 Claude Opus 4.8
  37. 29.4 Claude Sonnet 5
  38. 28.5 MiMo-V2.6 Flash
  39. 28.1 DeepSeek V4 Flash 0731
  40. 27.3 Step 5 Preview
  41. 26.8 DeepSeek V4 Pro 0813
  42. 26.6 DeepSeek V4.1 Flash
  43. 26.3 Muse Spark Open
  44. 25.8 Claude Sonnet 4.6
  45. 25.2 DeepSeek-V4-Pro
  46. 24.9 Kimi K3.1 Pro
  47. 24.9 Grok 4.5
  48. 24.8 GPT-5.6 Luna
  49. 23.7 Claude Opus 4.7
  50. 22.8 Gemini 3.7 Flash
  51. 21.8 GPT-5.5
  52. 21.8 GPT-6.1 Sol
  53. 21.6 Gemini 3.5 Flash
  54. 20.7 GPT-6 Astra
  55. 20.6 Ling-3.0-flash-Fin
  56. 20.6 MiniMax-M3
  57. 19.9 GPT-6 Luna
  58. 18.2 Mercury 3
  59. 18.0 Gemini 3.6 Flash
  60. 17.1 GPT-6.1 Terra
  61. 15.8 Qwen3.7-Max
  62. 15.4 Inkling-Small
  63. 15.3 Inkling
  64. 15.0 GPT-6 Sol
  65. 12.4 Kimi K2.6
  66. 11.8 Gemini 3 Flash
  67. 11.6 Nemotron 3.5 Super VL
  68. 11.5 MiMo-V2.5-Pro
  69. 10.5 Qwen3.7-Plus
  70. 9.29 Gemini 3.1 Pro
  71. 8.81 GPT-5.4 Mini
  72. 8.81 Beam 1.0
  73. 7.77 MiMo-V2.5
  74. 7.68 Qwen 3.6 Plus
  75. 7.25 Gemini 3.5 Flash-Lite
  76. 6.95 Grok 4.20
  77. 6.87 MiniMax-M2.7
  78. 6.18 Grok 4.3
  79. 5.86 Claude Haiku 4.5
  80. 4.97 GPT-5.4 Nano
  81. 4.48 Mistral Medium 3.5
  82. 3.11 Gemini 3.1 Flash-Lite
  83. 2.59 Nemotron 3.5 Lightning 30B-A3B
  84. 2.49 Mercury 2.5
  1. 1 Inkling 2 Small 65.8%
  2. 2 Gemini 4 Hydrogen 52.3%
  3. 3 Gemini 4 Carbon 51.8%
  4. 4 Claude Fable 5.1 50.8%
  5. 5 Qwen3.8 Max (0902) 49.7%
  6. 6 Gemini 4 Argon 49.2%
  7. 7 Claude Opus 5 47.1%
  8. 8 Claude Opus 5.5 44.5%
  9. 9 DeepSeek V4.1 Flash 1017 44.4%
  10. 10 GLM-5.3 43.3%
  11. 11 Muse Wonder Full 43%
  12. 12 Muse Spark 1.3 max 42.4%
  13. 13 Claude Sonnet 5.5 42%
  14. 14 Claude Fable 5 38.3%
  15. 15 Muse Ember 38.2%
  16. 16 Muse Wonder 36.3%
  17. 17 Grok 4.7 35.8%
  18. 18 Tencent Hy4 34.4%
  19. 19 Grok 4.6 33.7%
  20. 20 Kimi K3 33.4%
  21. 21 Muse Spark 1.2 32.8%
  22. 22 Tencent Hy4 preview 32.6%
  23. 23 MiMo-V2.6 Pro 32.4%
  24. 24 MiniMax-M3.1-Flash 32.1%
  25. 25 Gemini 3.8 Flash 32.1%
  26. 26 Qwen 3.8 Max 32.1%
  27. 27 Claude Fable 5.2 32.1%
  28. 28 Claude Haiku 5.5 31.9%
  29. 29 GLM-5.3-Flash 31.6%
  30. 30 GPT-5.6 Sol 31.1%
  31. 31 Ember-1 31.1%
  32. 32 Qwen 3.8 27B 30.9%
  33. 33 Mistral Large 4 30.7%
  34. 34 Mistral Large 4 Preview 30.4%
  35. 35 GPT-5.6 Terra 30%
  36. 36 Claude Opus 4.8 30%
  37. 37 Claude Sonnet 5 29.4%
  38. 38 MiMo-V2.6 Flash 28.5%
  39. 39 DeepSeek V4 Flash 0731 28.1%
  40. 40 Step 5 Preview 27.3%
  41. 41 DeepSeek V4 Pro 0813 26.8%
  42. 42 DeepSeek V4.1 Flash 26.6%
  43. 43 Muse Spark Open 26.3%
  44. 44 Claude Sonnet 4.6 25.8%
  45. 45 DeepSeek-V4-Pro 25.2%
  46. 46 Kimi K3.1 Pro 24.9%
  47. 47 Grok 4.5 24.9%
  48. 48 GPT-5.6 Luna 24.8%
  49. 49 Claude Opus 4.7 23.7%
  50. 50 Gemini 3.7 Flash 22.8%
  51. 51 GPT-5.5 21.8%
  52. 52 GPT-6.1 Sol 21.8%
  53. 53 Gemini 3.5 Flash 21.6%
  54. 54 GPT-6 Astra 20.7%
  55. 55 Ling-3.0-flash-Fin 20.6%
  56. 56 MiniMax-M3 20.6%
  57. 57 GPT-6 Luna 19.9%
  58. 58 Mercury 3 18.2%
  59. 59 Gemini 3.6 Flash 18%
  60. 60 GPT-6.1 Terra 17.1%
  61. 61 Qwen3.7-Max 15.8%
  62. 62 Inkling-Small 15.4%
  63. 63 Inkling 15.3%
  64. 64 GPT-6 Sol 15%
  65. 65 Kimi K2.6 12.4%
  66. 66 Gemini 3 Flash 11.8%
  67. 67 Nemotron 3.5 Super VL 11.6%
  68. 68 MiMo-V2.5-Pro 11.5%
  69. 69 Qwen3.7-Plus 10.5%
  70. 70 Gemini 3.1 Pro 9.29%
  71. 71 GPT-5.4 Mini 8.81%
  72. 72 Beam 1.0 8.81%
  73. 73 MiMo-V2.5 7.77%
  74. 74 Qwen 3.6 Plus 7.68%
  75. 75 Gemini 3.5 Flash-Lite 7.25%
  76. 76 Grok 4.20 reasoning 6.95%
  77. 77 MiniMax-M2.7 6.87%
  78. 78 Grok 4.3 6.18%
  79. 79 Claude Haiku 4.5 thinking 5.86%
  80. 80 GPT-5.4 Nano 4.97%
  81. 81 Mistral Medium 3.5 4.48%
  82. 82 Gemini 3.1 Flash-Lite 3.11%
  83. 83 Nemotron 3.5 Lightning 30B-A3B 2.59%
  84. 84 Mercury 2.5 2.49%

Sources: Labs’ model cards and official Tax Agent Bench leaderboard

#ModelScoreSettingDateSource
1 Gemini 4 Hydrogen 52.3% Nov 2026 Official leaderboard
2 Gemini 4 Carbon 51.8% Oct 2026 Official leaderboard
3 Claude Fable 5.1 50.8% Sep 2026 Official leaderboard
4 Qwen3.8 Max (0902) 49.7% Sep 2026 Official leaderboard
5 Gemini 4 Argon 49.2% Sep 2026 Official leaderboard
6 Claude Opus 5 47.1% Jul 2026 Official leaderboard
7 Claude Opus 5.5 44.5% Sep 2026 Official leaderboard
8 DeepSeek V4.1 Flash 1017 44.4% Oct 2026 Official leaderboard
9 GLM-5.3 43.3% Aug 2026 Official leaderboard
10 Muse Wonder Full 43% Oct 2026 Official leaderboard
11 Muse Spark 1.3 42.4% max Sep 2026 Official leaderboard
12 Claude Sonnet 5.5 42% Sep 2026 Official leaderboard
13 Inkling 2 Small 38.9% Nov 2026 Official leaderboard
14 Claude Fable 5 38.3% Jun 2026 Official leaderboard
15 Muse Ember 38.2% Nov 2026 Official leaderboard
16 Muse Wonder 36.3% Oct 2026 Official leaderboard
17 Grok 4.7 35.8% Sep 2026 Official leaderboard
18 Tencent Hy4 34.4% Nov 2026 Official leaderboard
19 Grok 4.6 33.7% Aug 2026 Official leaderboard
20 Kimi K3 33.4% Jul 2026 Official leaderboard
21 Muse Spark 1.2 32.8% Aug 2026 Official leaderboard
22 Tencent Hy4 preview 32.6% Aug 2026 Official leaderboard
23 MiMo-V2.6 Pro 32.4% Sep 2026 Official leaderboard
24 MiniMax-M3.1-Flash 32.1% Oct 2026 Official leaderboard
25 Gemini 3.8 Flash 32.1% Sep 2026 Official leaderboard
26 Qwen 3.8 Max 32.1% Aug 2026 Official leaderboard
27 Claude Fable 5.2 32.1% Nov 2026 Official leaderboard
28 Claude Haiku 5.5 31.9% Oct 2026 Official leaderboard
29 GLM-5.3-Flash 31.6% Aug 2026 Official leaderboard
30 GPT-5.6 Sol 31.1% Jul 2026 Official leaderboard
31 Ember-1 31.1% Sep 2026 Official leaderboard
32 Qwen 3.8 27B 30.9% Aug 2026 Official leaderboard
33 Mistral Large 4 30.7% Oct 2026 Official leaderboard
34 Mistral Large 4 Preview 30.4% Oct 2026 Official leaderboard
35 GPT-5.6 Terra 30% Jul 2026 Official leaderboard
36 Claude Opus 4.8 30% May 2026 Official leaderboard
37 Claude Sonnet 5 29.4% Jun 2026 Official leaderboard
38 MiMo-V2.6 Flash 28.5% Sep 2026 Official leaderboard
39 DeepSeek V4 Flash 0731 28.1% Jul 2026 Official leaderboard
40 Step 5 Preview 27.3% Sep 2026 Official leaderboard
41 DeepSeek V4 Pro 0813 26.8% Aug 2026 Official leaderboard
42 DeepSeek V4.1 Flash 26.6% Sep 2026 Official leaderboard
43 Muse Spark Open 26.3% Oct 2026 Official leaderboard
44 Claude Sonnet 4.6 25.8% Feb 2026 Official leaderboard
45 DeepSeek-V4-Pro 25.2% Apr 2026 Official leaderboard
46 Kimi K3.1 Pro 24.9% Oct 2026 Official leaderboard
47 Grok 4.5 24.9% Jul 2026 Official leaderboard
48 GPT-5.6 Luna 24.8% Jul 2026 Official leaderboard
49 Claude Opus 4.7 23.7% Apr 2026 Official leaderboard
50 Gemini 3.7 Flash 22.8% Aug 2026 Official leaderboard
51 GPT-5.5 21.8% Apr 2026 Official leaderboard
52 GPT-6.1 Sol 21.8% Sep 2026 Official leaderboard
53 Gemini 3.5 Flash 21.6% May 2026 Official leaderboard
54 GPT-6 Astra 20.7% Sep 2026 Official leaderboard
55 Ling-3.0-flash-Fin 20.6% Sep 2026 Official leaderboard
56 MiniMax-M3 20.6% Jun 2026 Official leaderboard
57 GPT-6 Luna 19.9% Sep 2026 Official leaderboard
58 Mercury 3 18.2% Oct 2026 Official leaderboard
59 Gemini 3.6 Flash 18% Jul 2026 Official leaderboard
60 GPT-6.1 Terra 17.1% Oct 2026 Official leaderboard
61 Qwen3.7-Max 15.8% May 2026 Official leaderboard
62 Inkling-Small 15.4% Jul 2026 Official leaderboard
63 Inkling 15.3% Jul 2026 Official leaderboard
64 GPT-6 Sol 15% Sep 2026 Official leaderboard
65 Kimi K2.6 12.4% Apr 2026 Official leaderboard
66 Gemini 3 Flash 11.8% Dec 2025 Official leaderboard
67 Nemotron 3.5 Super VL 11.6% Oct 2026 Official leaderboard
68 MiMo-V2.5-Pro 11.5% Apr 2026 Official leaderboard
69 Qwen3.7-Plus 10.5% Jun 2026 Official leaderboard
70 Gemini 3.1 Pro 9.29% Feb 2026 Official leaderboard
71 GPT-5.4 Mini 8.81% Mar 2026 Official leaderboard
72 Beam 1.0 8.81% Oct 2026 Official leaderboard
73 MiMo-V2.5 7.77% Apr 2026 Official leaderboard
74 Qwen 3.6 Plus 7.68% Mar 2026 Official leaderboard
75 Gemini 3.5 Flash-Lite 7.25% Jul 2026 Official leaderboard
76 Grok 4.20 6.95% reasoning Feb 2026 Official leaderboard
77 MiniMax-M2.7 6.87% Mar 2026 Official leaderboard
78 Grok 4.3 6.18% Apr 2026 Official leaderboard
79 Claude Haiku 4.5 5.86% thinking Oct 2025 Official leaderboard
80 GPT-5.4 Nano 4.97% Mar 2026 Official leaderboard
81 Mistral Medium 3.5 4.48% Apr 2026 Official leaderboard
82 Gemini 3.1 Flash-Lite 3.11% Mar 2026 Official leaderboard
83 Nemotron 3.5 Lightning 30B-A3B 2.59% Aug 2026 Official leaderboard
84 Mercury 2.5 2.49% Sep 2026 Official leaderboard
#ModelScoreSettingDateSource
1 Inkling 2 Small 65.8% Nov 2026 Lab model card
2 Gemini 4 Hydrogen 52.3% Nov 2026 Official leaderboard
3 Gemini 4 Carbon 51.8% Oct 2026 Official leaderboard
4 Claude Fable 5.1 50.8% Sep 2026 Official leaderboard
5 Qwen3.8 Max (0902) 49.7% Sep 2026 Official leaderboard
6 Gemini 4 Argon 49.2% Sep 2026 Official leaderboard
7 Claude Opus 5 47.1% Jul 2026 Official leaderboard
8 Claude Opus 5.5 44.5% Sep 2026 Official leaderboard
9 DeepSeek V4.1 Flash 1017 44.4% Oct 2026 Official leaderboard
10 GLM-5.3 43.3% Aug 2026 Official leaderboard
11 Muse Wonder Full 43% Oct 2026 Official leaderboard
12 Muse Spark 1.3 42.4% max Sep 2026 Official leaderboard
13 Claude Sonnet 5.5 42% Sep 2026 Official leaderboard
14 Claude Fable 5 38.3% Jun 2026 Official leaderboard
15 Muse Ember 38.2% Nov 2026 Official leaderboard
16 Muse Wonder 36.3% Oct 2026 Official leaderboard
17 Grok 4.7 35.8% Sep 2026 Official leaderboard
18 Tencent Hy4 34.4% Nov 2026 Official leaderboard
19 Grok 4.6 33.7% Aug 2026 Official leaderboard
20 Kimi K3 33.4% Jul 2026 Official leaderboard
21 Muse Spark 1.2 32.8% Aug 2026 Official leaderboard
22 Tencent Hy4 preview 32.6% Aug 2026 Official leaderboard
23 MiMo-V2.6 Pro 32.4% Sep 2026 Official leaderboard
24 MiniMax-M3.1-Flash 32.1% Oct 2026 Official leaderboard
25 Gemini 3.8 Flash 32.1% Sep 2026 Official leaderboard
26 Qwen 3.8 Max 32.1% Aug 2026 Official leaderboard
27 Claude Fable 5.2 32.1% Nov 2026 Official leaderboard
28 Claude Haiku 5.5 31.9% Oct 2026 Official leaderboard
29 GLM-5.3-Flash 31.6% Aug 2026 Official leaderboard
30 GPT-5.6 Sol 31.1% Jul 2026 Official leaderboard
31 Ember-1 31.1% Sep 2026 Official leaderboard
32 Qwen 3.8 27B 30.9% Aug 2026 Official leaderboard
33 Mistral Large 4 30.7% Oct 2026 Official leaderboard
34 Mistral Large 4 Preview 30.4% Oct 2026 Official leaderboard
35 GPT-5.6 Terra 30% Jul 2026 Official leaderboard
36 Claude Opus 4.8 30% May 2026 Official leaderboard
37 Claude Sonnet 5 29.4% Jun 2026 Official leaderboard
38 MiMo-V2.6 Flash 28.5% Sep 2026 Official leaderboard
39 DeepSeek V4 Flash 0731 28.1% Jul 2026 Official leaderboard
40 Step 5 Preview 27.3% Sep 2026 Official leaderboard
41 DeepSeek V4 Pro 0813 26.8% Aug 2026 Official leaderboard
42 DeepSeek V4.1 Flash 26.6% Sep 2026 Official leaderboard
43 Muse Spark Open 26.3% Oct 2026 Official leaderboard
44 Claude Sonnet 4.6 25.8% Feb 2026 Official leaderboard
45 DeepSeek-V4-Pro 25.2% Apr 2026 Official leaderboard
46 Kimi K3.1 Pro 24.9% Oct 2026 Official leaderboard
47 Grok 4.5 24.9% Jul 2026 Official leaderboard
48 GPT-5.6 Luna 24.8% Jul 2026 Official leaderboard
49 Claude Opus 4.7 23.7% Apr 2026 Official leaderboard
50 Gemini 3.7 Flash 22.8% Aug 2026 Official leaderboard
51 GPT-5.5 21.8% Apr 2026 Official leaderboard
52 GPT-6.1 Sol 21.8% Sep 2026 Official leaderboard
53 Gemini 3.5 Flash 21.6% May 2026 Official leaderboard
54 GPT-6 Astra 20.7% Sep 2026 Official leaderboard
55 Ling-3.0-flash-Fin 20.6% Sep 2026 Official leaderboard
56 MiniMax-M3 20.6% Jun 2026 Official leaderboard
57 GPT-6 Luna 19.9% Sep 2026 Official leaderboard
58 Mercury 3 18.2% Oct 2026 Official leaderboard
59 Gemini 3.6 Flash 18% Jul 2026 Official leaderboard
60 GPT-6.1 Terra 17.1% Oct 2026 Official leaderboard
61 Qwen3.7-Max 15.8% May 2026 Official leaderboard
62 Inkling-Small 15.4% Jul 2026 Official leaderboard
63 Inkling 15.3% Jul 2026 Official leaderboard
64 GPT-6 Sol 15% Sep 2026 Official leaderboard
65 Kimi K2.6 12.4% Apr 2026 Official leaderboard
66 Gemini 3 Flash 11.8% Dec 2025 Official leaderboard
67 Nemotron 3.5 Super VL 11.6% Oct 2026 Official leaderboard
68 MiMo-V2.5-Pro 11.5% Apr 2026 Official leaderboard
69 Qwen3.7-Plus 10.5% Jun 2026 Official leaderboard
70 Gemini 3.1 Pro 9.29% Feb 2026 Official leaderboard
71 GPT-5.4 Mini 8.81% Mar 2026 Official leaderboard
72 Beam 1.0 8.81% Oct 2026 Official leaderboard
73 MiMo-V2.5 7.77% Apr 2026 Official leaderboard
74 Qwen 3.6 Plus 7.68% Mar 2026 Official leaderboard
75 Gemini 3.5 Flash-Lite 7.25% Jul 2026 Official leaderboard
76 Grok 4.20 6.95% reasoning Feb 2026 Official leaderboard
77 MiniMax-M2.7 6.87% Mar 2026 Official leaderboard
78 Grok 4.3 6.18% Apr 2026 Official leaderboard
79 Claude Haiku 4.5 5.86% thinking Oct 2025 Official leaderboard
80 GPT-5.4 Nano 4.97% Mar 2026 Official leaderboard
81 Mistral Medium 3.5 4.48% Apr 2026 Official leaderboard
82 Gemini 3.1 Flash-Lite 3.11% Mar 2026 Official leaderboard
83 Nemotron 3.5 Lightning 30B-A3B 2.59% Aug 2026 Official leaderboard
84 Mercury 2.5 2.49% Sep 2026 Official leaderboard

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

Accuracy against latency

The Vals scatter chart with its x axis switched from Cost to Latency: mean time per test against accuracy, on a log scale. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

0% 20% 40% 60% 101001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Claude Haiku 4.5 (thinking): Accuracy 5.86%, latency (mean per test) 155 s Gemini 3 Flash: Accuracy 11.8%, latency (mean per test) 101 s Claude Sonnet 4.6: Accuracy 25.8%, latency (mean per test) 707 s Grok 4.20 (reasoning): Accuracy 6.95%, latency (mean per test) 89.3 s Gemini 3.1 Pro: Accuracy 9.29%, latency (mean per test) 101 s GPT-5.4 Mini: Accuracy 8.81%, latency (mean per test) 1,006 s GPT-5.4 Nano: Accuracy 4.97%, latency (mean per test) 220 s MiniMax-M2.7: Accuracy 6.87%, latency (mean per test) 303 s Qwen 3.6 Plus: Accuracy 7.68%, latency (mean per test) 378 s Claude Opus 4.7: Accuracy 23.7%, latency (mean per test) 279 s Kimi K2.6: Accuracy 12.4%, latency (mean per test) 924 s GPT-5.5: Accuracy 21.8%, latency (mean per test) 952 s MiMo-V2.5: Accuracy 7.77%, latency (mean per test) 501 s MiMo-V2.5-Pro: Accuracy 11.5%, latency (mean per test) 377 s DeepSeek-V4-Pro: Accuracy 25.2%, latency (mean per test) 1,245 s Mistral Medium 3.5: Accuracy 4.48%, latency (mean per test) 431 s Grok 4.3: Accuracy 6.18%, latency (mean per test) 97.5 s Gemini 3.5 Flash: Accuracy 21.6%, latency (mean per test) 233 s Qwen3.7-Max: Accuracy 15.8%, latency (mean per test) 210 s Claude Opus 4.8: Accuracy 30%, latency (mean per test) 479 s MiniMax-M3: Accuracy 20.6%, latency (mean per test) 315 s Qwen3.7-Plus: Accuracy 10.5%, latency (mean per test) 310 s Claude Fable 5: Accuracy 38.3%, latency (mean per test) 867 s Claude Sonnet 5: Accuracy 29.4%, latency (mean per test) 1,159 s Grok 4.5: Accuracy 24.9%, latency (mean per test) 284 s GPT-5.6 Luna: Accuracy 24.8%, latency (mean per test) 2,566 s GPT-5.6 Sol: Accuracy 31.1%, latency (mean per test) 2,681 s GPT-5.6 Terra: Accuracy 30%, latency (mean per test) 3,631 s Inkling: Accuracy 15.3%, latency (mean per test) 811 s Inkling-Small: Accuracy 15.4%, latency (mean per test) 385 s Kimi K3: Accuracy 33.4%, latency (mean per test) 2,415 s Gemini 3.6 Flash: Accuracy 18%, latency (mean per test) 105 s Claude Opus 5: Accuracy 47.1%, latency (mean per test) 1,008 s DeepSeek V4 Flash 0731: Accuracy 28.1%, latency (mean per test) 337 s Qwen 3.8 Max: Accuracy 32.1%, latency (mean per test) 2,436 s Muse Spark 1.2: Accuracy 32.8%, latency (mean per test) 120 s Nemotron 3.5 Lightning 30B-A3B: Accuracy 2.59%, latency (mean per test) 1,180 s Grok 4.6: Accuracy 33.7%, latency (mean per test) 658 s DeepSeek V4 Pro 0813: Accuracy 26.8%, latency (mean per test) 1,597 s GLM-5.3: Accuracy 43.3%, latency (mean per test) 3,053 s Qwen 3.8 27B: Accuracy 30.9%, latency (mean per test) 1,334 s GLM-5.3-Flash: Accuracy 31.6%, latency (mean per test) 1,297 s Tencent Hy4 preview: Accuracy 32.6%, latency (mean per test) 2,464 s Claude Fable 5.1: Accuracy 50.8%, latency (mean per test) 1,714 s Qwen3.8 Max (0902): Accuracy 49.7%, latency (mean per test) 5,231 s Gemini 3.8 Flash: Accuracy 32.1%, latency (mean per test) 178 s Muse Spark 1.3: Accuracy 37.9%, latency (mean per test) 505 s GPT-6 Astra: Accuracy 20.7%, latency (mean per test) 901 s Ling-3.0-flash-Fin: Accuracy 20.6%, latency (mean per test) 334 s Mercury 2.5: Accuracy 2.49%, latency (mean per test) 29.7 s DeepSeek V4.1 Flash: Accuracy 26.6%, latency (mean per test) 349 s Step 5 Preview: Accuracy 27.3%, latency (mean per test) 1,289 s Grok 4.7: Accuracy 35.8%, latency (mean per test) 996 s MiMo-V2.6 Flash: Accuracy 28.5%, latency (mean per test) 1,039 s MiMo-V2.6 Pro: Accuracy 32.4%, latency (mean per test) 1,548 s Claude Opus 5.5: Accuracy 44.5%, latency (mean per test) 4,013 s GPT-6 Luna: Accuracy 19.9%, latency (mean per test) 1,996 s GPT-6 Sol: Accuracy 15%, latency (mean per test) 1,052 s Ember-1: Accuracy 31.1%, latency (mean per test) 1,846 s Claude Sonnet 5.5: Accuracy 42%, latency (mean per test) 3,561 s GPT-6.1 Sol: Accuracy 21.8%, latency (mean per test) 1,214 s Mistral Large 4: Accuracy 30.7%, latency (mean per test) 1,023 s Mistral Large 4 Preview: Accuracy 30.4%, latency (mean per test) 3,438 s Claude Haiku 5.5: Accuracy 31.9%, latency (mean per test) 2,974 s Muse Wonder: Accuracy 36.3%, latency (mean per test) 493 s Muse Wonder Full: Accuracy 43%, latency (mean per test) 1,084 s Muse Spark Open: Accuracy 26.3%, latency (mean per test) 702 s DeepSeek V4.1 Flash 1017: Accuracy 44.4%, latency (mean per test) 1,264 s Mercury 3: Accuracy 18.2%, latency (mean per test) 269 s MiniMax-M3.1-Flash: Accuracy 32.1%, latency (mean per test) 1,963 s Kimi K3.1 Pro: Accuracy 24.9%, latency (mean per test) 3,519 s GPT-6.1 Terra: Accuracy 17.1%, latency (mean per test) 2,751 s Gemini 4 Carbon: Accuracy 51.8%, latency (mean per test) 2,086 s Tencent Hy4: Accuracy 34.4%, latency (mean per test) 3,808 s Inkling 2 Small: Accuracy 38.9%, latency (mean per test) 414 s Claude Fable 5.2: Accuracy 32.1%, latency (mean per test) 300 s Gemini 3.1 Flash-Lite: Accuracy 3.11%, latency (mean per test) 18.7 s Gemini 3.5 Flash-Lite: Accuracy 7.25%, latency (mean per test) 47.5 s Gemini 3.7 Flash: Accuracy 22.8%, latency (mean per test) 87.2 s Muse Spark 1.3 (max): Accuracy 42.4%, latency (mean per test) 249 s Gemini 4 Argon: Accuracy 49.2%, latency (mean per test) 986 s Beam 1.0: Accuracy 8.81%, latency (mean per test) 58 s Nemotron 3.5 Super VL: Accuracy 11.6%, latency (mean per test) 72.7 s Muse Ember: Accuracy 38.2%, latency (mean per test) 90.1 s Gemini 4 Hydrogen: Accuracy 52.3%, latency (mean per test) 1,095 s Gemini 4 HydrogenGemini 4 ArgonMuse Spark 1.3 (max)Muse EmberGemini 3.7 FlashNemotron 3.5 Super VLBeam 1.0Gemini 3.5 Flash-LiteGemini 3.1 Flash-Lite 0% 20% 40% 60% 101001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Claude Haiku 4.5 (thinking): Accuracy 5.86%, latency (mean per test) 155 s Gemini 3 Flash: Accuracy 11.8%, latency (mean per test) 101 s Claude Sonnet 4.6: Accuracy 25.8%, latency (mean per test) 707 s Grok 4.20 (reasoning): Accuracy 6.95%, latency (mean per test) 89.3 s Gemini 3.1 Pro: Accuracy 9.29%, latency (mean per test) 101 s GPT-5.4 Mini: Accuracy 8.81%, latency (mean per test) 1,006 s GPT-5.4 Nano: Accuracy 4.97%, latency (mean per test) 220 s MiniMax-M2.7: Accuracy 6.87%, latency (mean per test) 303 s Qwen 3.6 Plus: Accuracy 7.68%, latency (mean per test) 378 s Claude Opus 4.7: Accuracy 23.7%, latency (mean per test) 279 s Kimi K2.6: Accuracy 12.4%, latency (mean per test) 924 s GPT-5.5: Accuracy 21.8%, latency (mean per test) 952 s MiMo-V2.5: Accuracy 7.77%, latency (mean per test) 501 s MiMo-V2.5-Pro: Accuracy 11.5%, latency (mean per test) 377 s DeepSeek-V4-Pro: Accuracy 25.2%, latency (mean per test) 1,245 s Mistral Medium 3.5: Accuracy 4.48%, latency (mean per test) 431 s Grok 4.3: Accuracy 6.18%, latency (mean per test) 97.5 s Gemini 3.5 Flash: Accuracy 21.6%, latency (mean per test) 233 s Qwen3.7-Max: Accuracy 15.8%, latency (mean per test) 210 s Claude Opus 4.8: Accuracy 30%, latency (mean per test) 479 s MiniMax-M3: Accuracy 20.6%, latency (mean per test) 315 s Qwen3.7-Plus: Accuracy 10.5%, latency (mean per test) 310 s Claude Fable 5: Accuracy 38.3%, latency (mean per test) 867 s Claude Sonnet 5: Accuracy 29.4%, latency (mean per test) 1,159 s Grok 4.5: Accuracy 24.9%, latency (mean per test) 284 s GPT-5.6 Luna: Accuracy 24.8%, latency (mean per test) 2,566 s GPT-5.6 Sol: Accuracy 31.1%, latency (mean per test) 2,681 s GPT-5.6 Terra: Accuracy 30%, latency (mean per test) 3,631 s Inkling: Accuracy 15.3%, latency (mean per test) 811 s Inkling-Small: Accuracy 15.4%, latency (mean per test) 385 s Kimi K3: Accuracy 33.4%, latency (mean per test) 2,415 s Gemini 3.6 Flash: Accuracy 18%, latency (mean per test) 105 s Claude Opus 5: Accuracy 47.1%, latency (mean per test) 1,008 s DeepSeek V4 Flash 0731: Accuracy 28.1%, latency (mean per test) 337 s Qwen 3.8 Max: Accuracy 32.1%, latency (mean per test) 2,436 s Muse Spark 1.2: Accuracy 32.8%, latency (mean per test) 120 s Nemotron 3.5 Lightning 30B-A3B: Accuracy 2.59%, latency (mean per test) 1,180 s Grok 4.6: Accuracy 33.7%, latency (mean per test) 658 s DeepSeek V4 Pro 0813: Accuracy 26.8%, latency (mean per test) 1,597 s GLM-5.3: Accuracy 43.3%, latency (mean per test) 3,053 s Qwen 3.8 27B: Accuracy 30.9%, latency (mean per test) 1,334 s GLM-5.3-Flash: Accuracy 31.6%, latency (mean per test) 1,297 s Tencent Hy4 preview: Accuracy 32.6%, latency (mean per test) 2,464 s Claude Fable 5.1: Accuracy 50.8%, latency (mean per test) 1,714 s Qwen3.8 Max (0902): Accuracy 49.7%, latency (mean per test) 5,231 s Gemini 3.8 Flash: Accuracy 32.1%, latency (mean per test) 178 s Muse Spark 1.3: Accuracy 37.9%, latency (mean per test) 505 s GPT-6 Astra: Accuracy 20.7%, latency (mean per test) 901 s Ling-3.0-flash-Fin: Accuracy 20.6%, latency (mean per test) 334 s Mercury 2.5: Accuracy 2.49%, latency (mean per test) 29.7 s DeepSeek V4.1 Flash: Accuracy 26.6%, latency (mean per test) 349 s Step 5 Preview: Accuracy 27.3%, latency (mean per test) 1,289 s Grok 4.7: Accuracy 35.8%, latency (mean per test) 996 s MiMo-V2.6 Flash: Accuracy 28.5%, latency (mean per test) 1,039 s MiMo-V2.6 Pro: Accuracy 32.4%, latency (mean per test) 1,548 s Claude Opus 5.5: Accuracy 44.5%, latency (mean per test) 4,013 s GPT-6 Luna: Accuracy 19.9%, latency (mean per test) 1,996 s GPT-6 Sol: Accuracy 15%, latency (mean per test) 1,052 s Ember-1: Accuracy 31.1%, latency (mean per test) 1,846 s Claude Sonnet 5.5: Accuracy 42%, latency (mean per test) 3,561 s GPT-6.1 Sol: Accuracy 21.8%, latency (mean per test) 1,214 s Mistral Large 4: Accuracy 30.7%, latency (mean per test) 1,023 s Mistral Large 4 Preview: Accuracy 30.4%, latency (mean per test) 3,438 s Claude Haiku 5.5: Accuracy 31.9%, latency (mean per test) 2,974 s Muse Wonder: Accuracy 36.3%, latency (mean per test) 493 s Muse Wonder Full: Accuracy 43%, latency (mean per test) 1,084 s Muse Spark Open: Accuracy 26.3%, latency (mean per test) 702 s DeepSeek V4.1 Flash 1017: Accuracy 44.4%, latency (mean per test) 1,264 s Mercury 3: Accuracy 18.2%, latency (mean per test) 269 s MiniMax-M3.1-Flash: Accuracy 32.1%, latency (mean per test) 1,963 s Kimi K3.1 Pro: Accuracy 24.9%, latency (mean per test) 3,519 s GPT-6.1 Terra: Accuracy 17.1%, latency (mean per test) 2,751 s Gemini 4 Carbon: Accuracy 51.8%, latency (mean per test) 2,086 s Tencent Hy4: Accuracy 34.4%, latency (mean per test) 3,808 s Inkling 2 Small: Accuracy 38.9%, latency (mean per test) 414 s Claude Fable 5.2: Accuracy 32.1%, latency (mean per test) 300 s Gemini 3.1 Flash-Lite: Accuracy 3.11%, latency (mean per test) 18.7 s Gemini 3.5 Flash-Lite: Accuracy 7.25%, latency (mean per test) 47.5 s Gemini 3.7 Flash: Accuracy 22.8%, latency (mean per test) 87.2 s Muse Spark 1.3 (max): Accuracy 42.4%, latency (mean per test) 249 s Gemini 4 Argon: Accuracy 49.2%, latency (mean per test) 986 s Beam 1.0: Accuracy 8.81%, latency (mean per test) 58 s Nemotron 3.5 Super VL: Accuracy 11.6%, latency (mean per test) 72.7 s Muse Ember: Accuracy 38.2%, latency (mean per test) 90.1 s Gemini 4 Hydrogen: Accuracy 52.3%, latency (mean per test) 1,095 s Gemini 4 HydrogenGemini 4 ArgonMuse Spark 1.3 (max)Muse EmberGemini 3.7 FlashNemotron 3.5 Super VLBeam 1.0Gemini 3.5 Flash-LiteGemini 3.1 Flash-Lite
ModelSettingAccuracyLatency (mean per test)
Gemini 4 Hydrogen 52.3% 1,095 s
Gemini 4 Carbon 51.8% 2,086 s
Claude Fable 5.1 50.8% 1,714 s
Qwen3.8 Max (0902) 49.7% 5,231 s
Gemini 4 Argon 49.2% 985.63 s
Claude Opus 5 47.1% 1,008 s
Claude Opus 5.5 44.5% 4,013 s
DeepSeek V4.1 Flash 1017 44.4% 1,264 s
GLM-5.3 43.3% 3,053 s
Muse Wonder Full 43% 1,084 s
Muse Spark 1.3 max 42.4% 248.82 s
Claude Sonnet 5.5 42% 3,561 s
Inkling 2 Small 38.9% 414.04 s
Claude Fable 5 38.3% 866.63 s
Muse Ember 38.2% 90.13 s
Muse Spark 1.3 37.9% 505.37 s
Muse Wonder 36.3% 492.84 s
Grok 4.7 35.8% 996.47 s
Tencent Hy4 34.4% 3,808 s
Grok 4.6 33.7% 657.84 s
Kimi K3 33.4% 2,415 s
Muse Spark 1.2 32.8% 119.91 s
Tencent Hy4 preview 32.6% 2,464 s
MiMo-V2.6 Pro 32.4% 1,548 s
MiniMax-M3.1-Flash 32.1% 1,963 s
Gemini 3.8 Flash 32.1% 178.33 s
Claude Fable 5.2 32.1% 299.74 s
Qwen 3.8 Max 32.1% 2,436 s
Claude Haiku 5.5 31.9% 2,974 s
GLM-5.3-Flash 31.6% 1,297 s
GPT-5.6 Sol 31.1% 2,681 s
Ember-1 31.1% 1,846 s
Qwen 3.8 27B 30.9% 1,334 s
Mistral Large 4 30.7% 1,023 s
Mistral Large 4 Preview 30.4% 3,438 s
GPT-5.6 Terra 30% 3,631 s
Claude Opus 4.8 30% 479.2 s
Claude Sonnet 5 29.4% 1,159 s
MiMo-V2.6 Flash 28.5% 1,039 s
DeepSeek V4 Flash 0731 28.1% 337.34 s
Step 5 Preview 27.3% 1,289 s
DeepSeek V4 Pro 0813 26.8% 1,597 s
DeepSeek V4.1 Flash 26.6% 349.08 s
Muse Spark Open 26.3% 701.83 s
Claude Sonnet 4.6 25.8% 707.31 s
DeepSeek-V4-Pro 25.2% 1,245 s
Kimi K3.1 Pro 24.9% 3,519 s
Grok 4.5 24.9% 283.57 s
GPT-5.6 Luna 24.8% 2,566 s
Claude Opus 4.7 23.7% 279.04 s
Gemini 3.7 Flash 22.8% 87.24 s
GPT-5.5 21.8% 951.81 s
GPT-6.1 Sol 21.8% 1,214 s
Gemini 3.5 Flash 21.7% 233 s
GPT-6 Astra 20.7% 900.59 s
Ling-3.0-flash-Fin 20.6% 333.83 s
MiniMax-M3 20.6% 314.57 s
GPT-6 Luna 20% 1,996 s
Mercury 3 18.2% 269.27 s
Gemini 3.6 Flash 18% 105.32 s
GPT-6.1 Terra 17.1% 2,751 s
Qwen3.7-Max 15.8% 209.73 s
Inkling-Small 15.4% 385.25 s
Inkling 15.3% 810.56 s
GPT-6 Sol 15% 1,052 s
Kimi K2.6 12.4% 924.42 s
Gemini 3 Flash 11.8% 100.62 s
Nemotron 3.5 Super VL 11.6% 72.66 s
MiMo-V2.5-Pro 11.5% 376.51 s
Qwen3.7-Plus 10.5% 310.26 s
Gemini 3.1 Pro 9.3% 100.96 s
Beam 1.0 8.8% 57.99 s
GPT-5.4 Mini 8.8% 1,006 s
MiMo-V2.5 7.8% 500.83 s
Qwen 3.6 Plus 7.7% 378.16 s
Gemini 3.5 Flash-Lite 7.3% 47.46 s
Grok 4.20 reasoning 7% 89.26 s
MiniMax-M2.7 6.9% 302.85 s
Grok 4.3 6.2% 97.52 s
Claude Haiku 4.5 thinking 5.9% 155.02 s
GPT-5.4 Nano 5% 220.31 s
Mistral Medium 3.5 4.5% 430.67 s
Gemini 3.1 Flash-Lite 3.1% 18.74 s
Nemotron 3.5 Lightning 30B-A3B 2.6% 1,180 s
Mercury 2.5 2.5% 29.67 s

The best score, 52.3%, costs $3.87 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

Google DeepMind Xiaomi Corp Meta DeepSeek Thinking Machines Cost frontier One model at each effort level
0% 20% 40% 60% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Gemini 4 Hydrogen Gemini 4 Argon DeepSeek V4.1 Flash 1017 Inkling 2 Small Muse Ember MiMo-V2.5 Gemini 3.1 Flash-Lite 0% 20% 40% 60% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Gemini 4 Hydrogen DeepSeek V4.1 Flash 1017

Source: Official Tax Agent Bench leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
Gemini 4 Hydrogen 52.3% $3.87 307,129
Gemini 4 Carbon 51.8% $6.34 182,644
Claude Fable 5.1 50.8% $13 146,309
Qwen3.8 Max (0902) 49.7% $4.07 274,330
Gemini 4 Argon 49.2% $3.47 75,098
Claude Opus 5 47.1% $5.14 86,847
Claude Opus 5.5 44.5% $14 455,961
DeepSeek V4.1 Flash 1017 44.4% $0.46 277,522
GLM-5.3 43.3% $1.80 245,685
Muse Wonder Full 43% $2.15 51,650
Muse Spark 1.3 max 42.4% $0.39 31,585
Claude Sonnet 5.5 42% $8.73 559,739
Inkling 2 Small 38.9% $0.14 72,443
Claude Fable 5 38.3% $6.70 73,091
Muse Ember 38.2% $0.03 32,504
Muse Spark 1.3 37.9% $0.32 24,548
Muse Wonder 36.3% $1.88 45,444
Grok 4.7 35.8% $1.80 56,070
Tencent Hy4 34.4% $1.69 414,322
Grok 4.6 33.7% $0.98 39,758
Kimi K3 33.4% $2.79 77,697
Muse Spark 1.2 32.8% $0.27 19,763
Tencent Hy4 preview 32.6% $0.59 149,760
MiMo-V2.6 Pro 32.4% $0.13 64,880
MiniMax-M3.1-Flash 32.1% $0.32 364,312
Gemini 3.8 Flash 32.1% $0.86 36,335
Qwen 3.8 Max 32.1% $1.55 106,496
Claude Fable 5.2 32.1% $3.31 33,788
Claude Haiku 5.5 31.9% $1.16 741,776
GLM-5.3-Flash 31.6% $0.14 148,072
GPT-5.6 Sol 31.1% $7.51 188,472
Ember-1 31.1% $2.97 71,061
Qwen 3.8 27B 30.9% $0.82 187,156
Mistral Large 4 30.7% $0.50 76,008
Mistral Large 4 Preview 30.4% $1.39 139,787
GPT-5.6 Terra 30% $4.83 243,374
Claude Opus 4.8 30% $1.69 38,739
Claude Sonnet 5 29.4% $1.79 113,409
MiMo-V2.6 Flash 28.5% $0.05 62,862
DeepSeek V4 Flash 0731 28.1% $0.16 76,525
Step 5 Preview 27.3% $0.55 84,980
DeepSeek V4 Pro 0813 26.8% $0.73 109,118
DeepSeek V4.1 Flash 26.6% $0.13 80,656
Muse Spark Open 26.3% $0.88 18,505
Claude Sonnet 4.6 25.8% $1.19 38,465
DeepSeek-V4-Pro 25.2% $0.72 110,422
Kimi K3.1 Pro 24.9% $3.64 168,857
Grok 4.5 24.9% $0.47 16,998
GPT-5.6 Luna 24.8% $0.47 259,870
Claude Opus 4.7 23.7% $1.27 21,092
Gemini 3.7 Flash 22.8% $0.48 20,337
GPT-5.5 21.8% $4.07 66,972
GPT-6.1 Sol 21.8% $1.32 65,893
Gemini 3.5 Flash 21.6% $0.86 30,433
GPT-6 Astra 20.7% $5.80 62,346
Ling-3.0-flash-Fin 20.6% $0.03 43,263
MiniMax-M3 20.6% $0.16 24,976
GPT-6 Luna 19.9% $0.19 210,696
Mercury 3 18.2% $0.05 73,519
Gemini 3.6 Flash 18% $0.36 16,362
GPT-6.1 Terra 17.1% $2.16 234,123
Qwen3.7-Max 15.8% $0.44 18,741
Inkling-Small 15.4% $0.09 38,499
Inkling 15.3% $0.35 37,668
GPT-6 Sol 15% $2.28 97,209
Kimi K2.6 12.4% $0.57 45,053
Gemini 3 Flash 11.8% $0.10 16,306
Nemotron 3.5 Super VL 11.6% $0.05 30,297
MiMo-V2.5-Pro 11.5% $0.04 12,625
Qwen3.7-Plus 10.5% $0.09 22,236
Gemini 3.1 Pro 9.29% $0.30 9,861
GPT-5.4 Mini 8.81% $0.81 137,206
Beam 1.0 8.81% 20,848
MiMo-V2.5 7.77% $0.02 15,557
Qwen 3.6 Plus 7.68% $0.30 17,531
Gemini 3.5 Flash-Lite 7.25% $0.08 9,682
Grok 4.20 reasoning 6.95% $0.12 8,326
MiniMax-M2.7 6.87% $0.12 14,118
Grok 4.3 6.18% $0.12 9,232
Claude Haiku 4.5 thinking 5.86% $0.23 14,371
GPT-5.4 Nano 4.97% $0.07 23,086
Mistral Medium 3.5 4.48% $1.44 36,331
Gemini 3.1 Flash-Lite 3.11% $0.01 4,699
Nemotron 3.5 Lightning 30B-A3B 2.59% $0.03 116,422
Mercury 2.5 2.49% $0.02 8,007

3 models have held number one on the official leaderboard since its release in September 2026 3 models have held number one with claims included since its release in September 2026

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

Anthropic Google DeepMind
Own leaderboard 3 leaders
Gemini 4 Hydrogen 52.3%

Source: Official Tax Agent 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% 2026 2027 Near the ceiling (85%) Inkling 2 Small Gemini 4 Carbon Claude Opus 5 Claude Fable 5 Claude Opus 4.8 Claude Sonnet 4.6 Gemini 3 Flash 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Inkling 2 Small Gemini 4 Carbon

Sources: Official Tax Agent Bench leaderboard and labs’ model cards

One lab 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.

Thinking Machines
DateLaunch postScoreAs the post writes it
Nov 5, 2026 Inkling 2 Small 65.8 Tax Agent Bench, 65.8%

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
Inkling 2 Small 65.8% Tax Agent Bench, 65.8% 38.9% Leaderboard +27 pts

More in Law, finance and medicine

All 179 in Law, finance and medicine
  • MedXpertQA-MM In use Models answer expert-level medical exam questions that combine images with clinical records and test results. Scoring is accuracy. Qwen 3.8 Max 80.4%
  • HealthBench Professional In use 525 physician-authored tasks from real clinician chats (care consult, documentation, research) with physician-adjudicated rubrics. Claude Sonnet 5.5 69.2%
  • HealthBench In use 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. GPT-5 67.2%
  • FinSearchComp In use Agents search the web to fetch live financial data, look up history, and run multi-step investigations, across 635 expert-written questions. Scored by answer accuracy. Ling-3.1-flash 81.7%
  • KBL In use Models answer Korean legal knowledge and reasoning tasks and Korean bar exam questions. It has 7 knowledge tasks, 4 reasoning tasks and 2,510 bar exam items. Solar Pro 4 77.5%
  • KorMedMCQA In use Models answer 7,469 multiple-choice questions from Korean licensing exams for doctors, nurses, pharmacists and dentists, held from 2012 to 2024. Solar Pro 4 93.2%