Benchmark Almanac

Tax Agent Bench

Gemini 4 Carbon 51.8% leads since Oct 2026 Not in model cards

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 2026-10-29.

Made by Vals AI Released Sep 2026 Finance Public tasks plus a private test set Last model added Oct 2026 Scores refreshed 2 Nov 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Graded by a model A language model scores each response, usually against a rubric.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
vals-ai/tax-agent-bench 5 stars Leaderboard
79
Models scored on its own leaderboard

Gemini 4 Carbon leads with 51.8%

Best score per model from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

79 models. Scroll sideways for the rest.

  1. 51.8 Gemini 4 Carbon
  2. 49.7 Qwen3.8 Max (0902)
  3. 49.2 Claude Fable 5.1
  4. 45.6 Claude Opus 5.5
  5. 45.5 Claude Opus 5
  6. 44.6 Gemini 4 Argon
  7. 44.4 DeepSeek V4.1 Flash 1017
  8. 43.0 Muse Wonder Full
  9. 42.4 Muse Spark 1.3
  10. 42.0 Claude Sonnet 5.5
  11. 39.7 GLM-5.3
  12. 38.3 Claude Fable 5
  13. 36.3 Muse Wonder
  14. 34.2 Grok 4.7
  15. 33.4 Kimi K3
  16. 33.1 Grok 4.6
  17. 32.8 Muse Spark 1.2
  18. 32.6 Tencent Hy4 preview
  19. 32.4 MiMo-V2.6 Pro
  20. 32.1 MiniMax-M3.1-Flash
  21. 32.1 Gemini 3.8 Flash
  22. 32.1 Qwen 3.8 Max
  23. 31.9 Claude Haiku 5.5
  24. 31.6 GLM-5.3-Flash
  25. 31.1 GPT-5.6 Sol
  26. 31.1 Ember-1
  27. 30.9 Qwen 3.8 27B
  28. 30.7 Mistral Large 4 (Le Chonk, open weights)
  29. 30.4 Mistral Large 4
  30. 30.0 GPT-5.6 Terra
  31. 30.0 Claude Opus 4.8
  32. 29.4 Claude Sonnet 5
  33. 28.5 MiMo-V2.6 Flash
  34. 28.1 DeepSeek V4 Flash 0731
  35. 27.3 Step 5 Preview
  36. 26.8 DeepSeek V4 Pro 0813
  37. 26.6 DeepSeek V4.1 Flash
  38. 26.3 Muse Spark Open
  39. 25.8 Claude Sonnet 4.6
  40. 25.2 DeepSeek-V4-Pro
  41. 24.9 Kimi K3.1 Pro
  42. 24.9 Grok 4.5
  43. 24.8 GPT-5.6 Luna
  44. 23.7 Claude Opus 4.7
  45. 22.8 Gemini 3.7 Flash
  46. 21.8 GPT-5.5
  47. 21.8 GPT-6.1 Sol
  48. 21.6 Gemini 3.5 Flash
  49. 20.7 GPT-6 Astra
  50. 20.6 Ling-3.0-flash-Fin
  51. 20.6 MiniMax-M3
  52. 19.9 GPT-6 Luna
  53. 18.2 Mercury 3
  54. 18.0 Gemini 3.6 Flash
  55. 17.1 GPT-6.1 Terra
  56. 15.8 Qwen3.7-Max
  57. 15.4 Inkling-Small
  58. 15.3 Inkling
  59. 15.0 GPT-6 Sol
  60. 12.4 Kimi K2.6
  61. 11.8 Gemini 3 Flash
  62. 11.6 Nemotron 3.5 Super VL
  63. 11.5 MiMo-V2.5-Pro
  64. 10.5 Qwen3.7-Plus
  65. 9.29 Gemini 3.1 Pro
  66. 8.81 GPT-5.4 Mini
  67. 8.81 Beam 1.0
  68. 7.77 MiMo-V2.5
  69. 7.68 Qwen 3.6 Plus
  70. 7.25 Gemini 3.5 Flash-Lite
  71. 6.95 Grok 4.20
  72. 6.87 MiniMax-M2.7
  73. 6.18 Grok 4.3
  74. 5.86 Claude Haiku 4.5
  75. 4.97 GPT-5.4 Nano
  76. 4.48 Mistral Medium 3.5
  77. 3.11 Gemini 3.1 Flash-Lite
  78. 2.59 Nemotron 3.5 Lightning 30B-A3B
  79. 2.49 Mercury 2.5
  1. 1 Gemini 4 Carbon 51.8%
  2. 2 Qwen3.8 Max (0902) 49.7%
  3. 3 Claude Fable 5.1 49.2%
  4. 4 Claude Opus 5.5 45.6%
  5. 5 Claude Opus 5 45.5%
  6. 6 Gemini 4 Argon 44.6%
  7. 7 DeepSeek V4.1 Flash 1017 44.4%
  8. 8 Muse Wonder Full 43%
  9. 9 Muse Spark 1.3 max 42.4%
  10. 10 Claude Sonnet 5.5 42%
  11. 11 GLM-5.3 39.7%
  12. 12 Claude Fable 5 38.3%
  13. 13 Muse Wonder 36.3%
  14. 14 Grok 4.7 34.2%
  15. 15 Kimi K3 33.4%
  16. 16 Grok 4.6 33.1%
  17. 17 Muse Spark 1.2 32.8%
  18. 18 Tencent Hy4 preview 32.6%
  19. 19 MiMo-V2.6 Pro 32.4%
  20. 20 MiniMax-M3.1-Flash 32.1%
  21. 21 Gemini 3.8 Flash 32.1%
  22. 22 Qwen 3.8 Max 32.1%
  23. 23 Claude Haiku 5.5 31.9%
  24. 24 GLM-5.3-Flash 31.6%
  25. 25 GPT-5.6 Sol 31.1%
  26. 26 Ember-1 31.1%
  27. 27 Qwen 3.8 27B 30.9%
  28. 28 Mistral Large 4 (Le Chonk, open weights) 30.7%
  29. 29 Mistral Large 4 30.4%
  30. 30 GPT-5.6 Terra 30%
  31. 31 Claude Opus 4.8 30%
  32. 32 Claude Sonnet 5 29.4%
  33. 33 MiMo-V2.6 Flash 28.5%
  34. 34 DeepSeek V4 Flash 0731 28.1%
  35. 35 Step 5 Preview 27.3%
  36. 36 DeepSeek V4 Pro 0813 26.8%
  37. 37 DeepSeek V4.1 Flash 26.6%
  38. 38 Muse Spark Open 26.3%
  39. 39 Claude Sonnet 4.6 25.8%
  40. 40 DeepSeek-V4-Pro 25.2%
  41. 41 Kimi K3.1 Pro 24.9%
  42. 42 Grok 4.5 24.9%
  43. 43 GPT-5.6 Luna 24.8%
  44. 44 Claude Opus 4.7 23.7%
  45. 45 Gemini 3.7 Flash 22.8%
  46. 46 GPT-5.5 21.8%
  47. 47 GPT-6.1 Sol 21.8%
  48. 48 Gemini 3.5 Flash 21.6%
  49. 49 GPT-6 Astra 20.7%
  50. 50 Ling-3.0-flash-Fin 20.6%
  51. 51 MiniMax-M3 20.6%
  52. 52 GPT-6 Luna 19.9%
  53. 53 Mercury 3 18.2%
  54. 54 Gemini 3.6 Flash 18%
  55. 55 GPT-6.1 Terra 17.1%
  56. 56 Qwen3.7-Max 15.8%
  57. 57 Inkling-Small 15.4%
  58. 58 Inkling 15.3%
  59. 59 GPT-6 Sol 15%
  60. 60 Kimi K2.6 12.4%
  61. 61 Gemini 3 Flash 11.8%
  62. 62 Nemotron 3.5 Super VL 11.6%
  63. 63 MiMo-V2.5-Pro 11.5%
  64. 64 Qwen3.7-Plus 10.5%
  65. 65 Gemini 3.1 Pro 9.29%
  66. 66 GPT-5.4 Mini 8.81%
  67. 67 Beam 1.0 8.81%
  68. 68 MiMo-V2.5 7.77%
  69. 69 Qwen 3.6 Plus 7.68%
  70. 70 Gemini 3.5 Flash-Lite 7.25%
  71. 71 Grok 4.20 6.95%
  72. 72 MiniMax-M2.7 6.87%
  73. 73 Grok 4.3 6.18%
  74. 74 Claude Haiku 4.5 thinking 5.86%
  75. 75 GPT-5.4 Nano 4.97%
  76. 76 Mistral Medium 3.5 4.48%
  77. 77 Gemini 3.1 Flash-Lite 3.11%
  78. 78 Nemotron 3.5 Lightning 30B-A3B 2.59%
  79. 79 Mercury 2.5 2.49%
#ModelScoreSettingDateSource
1 Gemini 4 Carbon 51.8% Oct 2026 Official leaderboard
2 Qwen3.8 Max (0902) 49.7% Sep 2026 Official leaderboard
3 Claude Fable 5.1 49.2% Sep 2026 Official leaderboard
4 Claude Opus 5.5 45.6% Sep 2026 Official leaderboard
5 Claude Opus 5 45.5% Jul 2026 Official leaderboard
6 Gemini 4 Argon 44.6% Sep 2026 Official leaderboard
7 DeepSeek V4.1 Flash 1017 44.4% Oct 2026 Official leaderboard
8 Muse Wonder Full 43% Oct 2026 Official leaderboard
9 Muse Spark 1.3 42.4% max Sep 2026 Official leaderboard
10 Claude Sonnet 5.5 42% Sep 2026 Official leaderboard
11 GLM-5.3 39.7% Aug 2026 Official leaderboard
12 Claude Fable 5 38.3% Jun 2026 Official leaderboard
13 Muse Wonder 36.3% Oct 2026 Official leaderboard
14 Grok 4.7 34.2% Sep 2026 Official leaderboard
15 Kimi K3 33.4% Jul 2026 Official leaderboard
16 Grok 4.6 33.1% Aug 2026 Official leaderboard
17 Muse Spark 1.2 32.8% Aug 2026 Official leaderboard
18 Tencent Hy4 preview 32.6% Aug 2026 Official leaderboard
19 MiMo-V2.6 Pro 32.4% Sep 2026 Official leaderboard
20 MiniMax-M3.1-Flash 32.1% Oct 2026 Official leaderboard
21 Gemini 3.8 Flash 32.1% Sep 2026 Official leaderboard
22 Qwen 3.8 Max 32.1% Aug 2026 Official leaderboard
23 Claude Haiku 5.5 31.9% Oct 2026 Official leaderboard
24 GLM-5.3-Flash 31.6% Aug 2026 Official leaderboard
25 GPT-5.6 Sol 31.1% Jul 2026 Official leaderboard
26 Ember-1 31.1% Sep 2026 Official leaderboard
27 Qwen 3.8 27B 30.9% Aug 2026 Official leaderboard
28 Mistral Large 4 (Le Chonk, open weights) 30.7% Oct 2026 Official leaderboard
29 Mistral Large 4 30.4% Oct 2026 Official leaderboard
30 GPT-5.6 Terra 30% Jul 2026 Official leaderboard
31 Claude Opus 4.8 30% May 2026 Official leaderboard
32 Claude Sonnet 5 29.4% Jun 2026 Official leaderboard
33 MiMo-V2.6 Flash 28.5% Sep 2026 Official leaderboard
34 DeepSeek V4 Flash 0731 28.1% Jul 2026 Official leaderboard
35 Step 5 Preview 27.3% Sep 2026 Official leaderboard
36 DeepSeek V4 Pro 0813 26.8% Aug 2026 Official leaderboard
37 DeepSeek V4.1 Flash 26.6% Sep 2026 Official leaderboard
38 Muse Spark Open 26.3% Oct 2026 Official leaderboard
39 Claude Sonnet 4.6 25.8% Feb 2026 Official leaderboard
40 DeepSeek-V4-Pro 25.2% Apr 2026 Official leaderboard
41 Kimi K3.1 Pro 24.9% Oct 2026 Official leaderboard
42 Grok 4.5 24.9% Jul 2026 Official leaderboard
43 GPT-5.6 Luna 24.8% Jul 2026 Official leaderboard
44 Claude Opus 4.7 23.7% Apr 2026 Official leaderboard
45 Gemini 3.7 Flash 22.8% Aug 2026 Official leaderboard
46 GPT-5.5 21.8% Apr 2026 Official leaderboard
47 GPT-6.1 Sol 21.8% Sep 2026 Official leaderboard
48 Gemini 3.5 Flash 21.6% May 2026 Official leaderboard
49 GPT-6 Astra 20.7% Sep 2026 Official leaderboard
50 Ling-3.0-flash-Fin 20.6% Sep 2026 Official leaderboard
51 MiniMax-M3 20.6% Jun 2026 Official leaderboard
52 GPT-6 Luna 19.9% Sep 2026 Official leaderboard
53 Mercury 3 18.2% Oct 2026 Official leaderboard
54 Gemini 3.6 Flash 18% Jul 2026 Official leaderboard
55 GPT-6.1 Terra 17.1% Oct 2026 Official leaderboard
56 Qwen3.7-Max 15.8% May 2026 Official leaderboard
57 Inkling-Small 15.4% Jul 2026 Official leaderboard
58 Inkling 15.3% Jul 2026 Official leaderboard
59 GPT-6 Sol 15% Sep 2026 Official leaderboard
60 Kimi K2.6 12.4% Apr 2026 Official leaderboard
61 Gemini 3 Flash 11.8% Dec 2025 Official leaderboard
62 Nemotron 3.5 Super VL 11.6% Oct 2026 Official leaderboard
63 MiMo-V2.5-Pro 11.5% Apr 2026 Official leaderboard
64 Qwen3.7-Plus 10.5% Jun 2026 Official leaderboard
65 Gemini 3.1 Pro 9.29% Feb 2026 Official leaderboard
66 GPT-5.4 Mini 8.81% Mar 2026 Official leaderboard
67 Beam 1.0 8.81% Oct 2026 Official leaderboard
68 MiMo-V2.5 7.77% Apr 2026 Official leaderboard
69 Qwen 3.6 Plus 7.68% Mar 2026 Official leaderboard
70 Gemini 3.5 Flash-Lite 7.25% Jul 2026 Official leaderboard
71 Grok 4.20 6.95% Feb 2026 Official leaderboard
72 MiniMax-M2.7 6.87% Mar 2026 Official leaderboard
73 Grok 4.3 6.18% Apr 2026 Official leaderboard
74 Claude Haiku 4.5 5.86% thinking Oct 2025 Official leaderboard
75 GPT-5.4 Nano 4.97% Mar 2026 Official leaderboard
76 Mistral Medium 3.5 4.48% Apr 2026 Official leaderboard
77 Gemini 3.1 Flash-Lite 3.11% Mar 2026 Official leaderboard
78 Nemotron 3.5 Lightning 30B-A3B 2.59% Aug 2026 Official leaderboard
79 Mercury 2.5 2.49% Sep 2026 Official leaderboard

The best score, 51.8%, costs $6.34 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 Qwen Ant Group DeepSeek 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 Carbon Gemini 4 Argon DeepSeek V4.1 Flash 1017 Muse Spark 1.3 Muse Spark 1.2 MiMo-V2.6 Flash Ling-3.0-flash-Fin 0% 20% 40% 60% $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale Gemini 4 Carbon Gemini 4 Argon
ModelEffortHarnessScoreCost per taskOutput tokens
Gemini 4 Carbon 51.8% $6.34
Qwen3.8 Max (0902) 49.7% $4.07
Claude Fable 5.1 49.2% $13
Claude Opus 5.5 45.6% $15
Claude Opus 5 45.5% $5.14
Gemini 4 Argon 44.6% $3.47
DeepSeek V4.1 Flash 1017 44.4% $0.46
Muse Wonder Full 43% $2.15
Muse Spark 1.3 max 42.4% $0.39
Claude Sonnet 5.5 42% $9.25
GLM-5.3 39.7% $1.80
Claude Fable 5 38.3% $6.70
Muse Wonder 36.3% $1.88
Muse Spark 1.3 36% $0.32
Grok 4.7 34.2% $1.80
Kimi K3 33.4% $2.79
Grok 4.6 33.1% $0.98
Muse Spark 1.2 32.8% $0.27
Tencent Hy4 preview 32.6% $0.59
MiMo-V2.6 Pro 32.4% $0.13
MiniMax-M3.1-Flash 32.1% $0.32
Gemini 3.8 Flash 32.1% $0.86
Qwen 3.8 Max 32.1% $1.55
Claude Haiku 5.5 31.9% $1.16
GLM-5.3-Flash 31.6% $0.14
GPT-5.6 Sol 31.1% $7.51
Ember-1 31.1% $2.97
Qwen 3.8 27B 30.9% $0.82
Mistral Large 4 (Le Chonk, open weights) 30.7% $0.50
Mistral Large 4 30.4% $1.39
GPT-5.6 Terra 30% $4.83
Claude Opus 4.8 30% $1.69
Claude Sonnet 5 29.4% $1.79
MiMo-V2.6 Flash 28.5% $0.05
DeepSeek V4 Flash 0731 28.1% $0.16
Step 5 Preview 27.3% $0.55
DeepSeek V4 Pro 0813 26.8% $0.73
DeepSeek V4.1 Flash 26.6% $0.13
Muse Spark Open 26.3% $0.88
Claude Sonnet 4.6 25.8% $1.19
DeepSeek-V4-Pro 25.2% $0.72
Kimi K3.1 Pro 24.9% $3.64
Grok 4.5 24.9% $0.47
GPT-5.6 Luna 24.8% $0.47
Claude Opus 4.7 23.7% $1.27
Gemini 3.7 Flash 22.8% $0.48
GPT-5.5 21.8% $4.07
GPT-6.1 Sol 21.8% $1.32
Gemini 3.5 Flash 21.6% $0.86
GPT-6 Astra 20.7% $5.80
Ling-3.0-flash-Fin 20.6% $0.03
MiniMax-M3 20.6% $0.16
GPT-6 Luna 19.9% $0.19
Mercury 3 18.2% $0.05
Gemini 3.6 Flash 18% $0.36
GPT-6.1 Terra 17.1% $2.16
Qwen3.7-Max 15.8% $0.44
Inkling-Small 15.4% $0.09
Inkling 15.3% $0.35
GPT-6 Sol 15% $2.28
Kimi K2.6 12.4% $0.57
Gemini 3 Flash 11.8% $0.10
Nemotron 3.5 Super VL 11.6% $0.05
MiMo-V2.5-Pro 11.5% $0.04
Qwen3.7-Plus 10.5% $0.09
Gemini 3.1 Pro 9.29% $0.30
GPT-5.4 Mini 8.81% $0.81
Beam 1.0 8.81%
MiMo-V2.5 7.77% $0.02
Qwen 3.6 Plus 7.68% $0.30
Gemini 3.5 Flash-Lite 7.25% $0.08
Grok 4.20 6.95% $0.12
MiniMax-M2.7 6.87% $0.12
Grok 4.3 6.18% $0.12
Claude Haiku 4.5 thinking 5.86% $0.23
GPT-5.4 Nano 4.97% $0.07
Mistral Medium 3.5 4.48% $1.44
Gemini 3.1 Flash-Lite 3.11% $0.01
Nemotron 3.5 Lightning 30B-A3B 2.59% $0.03
Mercury 2.5 2.49% $0.02

2 models have held number one since its release in Sep 2026

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

Qwen Google DeepMind
Own leaderboard 2 leaders
Gemini 4 Carbon 51.8%

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 Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Gemini 4 Carbon Claude Opus 5 Claude Fable 5 Claude Opus 4.8 Claude Sonnet 4.6 Gemini 3 Flash Claude Haiku 4.5 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Gemini 4 Carbon Claude Fable 5

No lab model card reports it yet

We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark.

More in Law, finance and medicine

All 140 in Law, finance and medicine
  • 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%
  • 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 In use 5,000 realistic health conversations graded against physician-written rubrics. A length-adjusted variant penalizes verbosity. GPT-5 67.2%
  • 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%
  • PRBench In use Model answers 1,100 expert-written law and finance questions. LLM judge grades against weighted expert rubrics of 10-30 criteria each. Muse Spark 1.3 61.6%