Vals AI

Tax Agent Bench

Agentic research on US tax questions, about 60% federal corporate income tax. The rest covers partnership, state and local, transfer pricing and employment tax. Agents use tools and must answer with citations.

Best reported

49.7

Qwen 3.8 Max 0902Qwen

accuracy % (all-pass x citation factor)

Official board, 29 Oct 2026

About this benchmark

Size
193 test questions
Public split
A public split exists (it includes example P-004), but its size is not stated
Grading
All-pass on expert rubric checks, scaled by citation quality: score = all-pass x (0.7 + 0.3 x citation quality). Weighted Pass is the secondary metric. An LLM judge grades.
Licence
proprietary
Submission
Not stated
Products allowed
Not stated
Latest version
2026-10-29
What it does not cover
US only. No personal tax filing and no non-US jurisdictions.

Results

10 models
Leaderboard

Each score shows who reported it: the benchmark's authors (quoted from their paper), the benchmark's official board, or Moneda (run with the benchmark's own grader and rules). We do not copy other aggregators.

Best value: Qwen 3.8 Max 0902, 49.7 at $4.07 per question (cheapest within 5% of the top score)

Tax Agent Bench: score vs cost

Score against cost per test. Up and to the left is better.

Updated 29 Oct 2026

Y: accuracy % (all-pass x citation factor)X: Cost per test (USD)Showing 10 models
Tax Agent Bench: score vs cost10 models. Frontier: Muse Spark 1.3 Max, DeepSeek V4.1 Flash 1017, Gemini 4 Argon, Qwen 3.8 Max 0902.4244464850$0$5$10$15EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Qwen 3.8 Max 0902: 49.7 accuracy % (all-pass x citation factor), $4.07 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Qwen 3.8 Max 0902Claude Fable 5.1: 49.2 accuracy % (all-pass x citation factor), $13.18 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Fable 5.1Gemini 4 Carbon: 48.9 accuracy % (all-pass x citation factor), $6.34 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 CarbonClaude Opus 5.5: 45.6 accuracy % (all-pass x citation factor), $15.09 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Opus 5.5Claude Opus 5: 45.5 accuracy % (all-pass x citation factor), $5.14 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Opus 5Gemini 4 Argon: 44.6 accuracy % (all-pass x citation factor), $3.47 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 ArgonDeepSeek V4.1 Flash 1017: 44.4 accuracy % (all-pass x citation factor), $0.46 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)DeepSeek V4.1 Flash 1017Wonder Full: 43 accuracy % (all-pass x citation factor), $2.15 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Wonder FullMuse Spark 1.3 Max: 42.4 accuracy % (all-pass x citation factor), $0.39 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Muse Spark 1.3 MaxClaude Sonnet 5.5: 42.0 accuracy % (all-pass x citation factor), $9.25 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Sonnet 5.5Tax Agent Bench: score vs cost10 models. Frontier: Muse Spark 1.3 Max, DeepSeek V4.1 Flash 1017, Gemini 4 Argon, Qwen 3.8 Max 0902.4244464850$0.5$1$2$5$10$20EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Qwen 3.8 Max 0902: 49.7 accuracy % (all-pass x citation factor), $4.07 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Qwen 3.8 Max 0902Claude Fable 5.1: 49.2 accuracy % (all-pass x citation factor), $13.18 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Fable 5.1Gemini 4 Carbon: 48.9 accuracy % (all-pass x citation factor), $6.34 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 CarbonClaude Opus 5.5: 45.6 accuracy % (all-pass x citation factor), $15.09 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Opus 5.5Claude Opus 5: 45.5 accuracy % (all-pass x citation factor), $5.14 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Opus 5Gemini 4 Argon: 44.6 accuracy % (all-pass x citation factor), $3.47 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 ArgonDeepSeek V4.1 Flash 1017: 44.4 accuracy % (all-pass x citation factor), $0.46 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)DeepSeek V4.1 Flash 1017Wonder Full: 43 accuracy % (all-pass x citation factor), $2.15 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Wonder FullMuse Spark 1.3 Max: 42.4 accuracy % (all-pass x citation factor), $0.39 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Muse Spark 1.3 MaxClaude Sonnet 5.5: 42.0 accuracy % (all-pass x citation factor), $9.25 per test (Vals.ai Tax Agent Bench leaderboard, Cost / Test, 29 Oct 2026)Claude Sonnet 5.5

Tax Agent Bench 2026-10-29

accuracy % (all-pass x citation factor), higher is better. 10 models.

Updated 29 Oct 2026

  1. Qwen 3.8 Max 090249.7
  2. Claude Fable 5.149.2
  3. Gemini 4 Carbon48.9
  4. Claude Opus 5.545.6
  5. Claude Opus 545.5
  6. Gemini 4 Argon44.6
  7. DeepSeek V4.1 Flash 101744.4
  8. Wonder Full43
  9. Muse Spark 1.3 Max42.4
  10. Claude Sonnet 5.542.0

Tax Agent Bench: cost

cost per test item (USD), lower is better. 10 models.

Updated 29 Oct 2026

  1. Muse Spark 1.3 Max0.390
  2. DeepSeek V4.1 Flash 10170.460
  3. Wonder Full2.1
  4. Gemini 4 Argon3.5
  5. Qwen 3.8 Max 09024.1
  6. Claude Opus 55.1
  7. Gemini 4 Carbon6.3
  8. Claude Sonnet 5.59.3
  9. Claude Fable 5.113.2
  10. Claude Opus 5.515.1

Tax Agent Bench 2026-10-29

accuracy % (all-pass x citation factor), higher is better. 10 models.

Updated 29 Oct 2026

  1. 1Qwen 3.8 Max 090249.7
  2. 2Claude Fable 5.149.2
  3. 3Gemini 4 Carbon48.9
  4. 4Claude Opus 5.545.6
  5. 5Claude Opus 545.5
  6. 6Gemini 4 Argon44.6
  7. 7DeepSeek V4.1 Flash 101744.4
  8. 8Wonder Full43
  9. 9Muse Spark 1.3 Max42.4
  10. 10Claude Sonnet 5.542.0
Tax Agent Bench leaderboard (accuracy % (all-pass x citation factor), higher is better)
1Qwen 3.8 Max 0902Qwen49.7test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
2Claude Fable 5.1Anthropic49.2test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
3Gemini 4 CarbonGoogle48.9test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
4Claude Opus 5.5Anthropic45.6test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
5Claude Opus 5Anthropic45.5test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
6Gemini 4 ArgonGoogle44.6test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
7DeepSeek V4.1 Flash 1017DeepSeek44.4test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
8Wonder FullNot stated43test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
9Muse Spark 1.3 MaxMeta42.4test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
10Claude Sonnet 5.5Anthropic42.0test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026

Last checked 1 Nov 2026. Something out of date or wrong? Send a correction.