Vals Tax Agent Bench (Tax Agent Bench)
We show this table for reference; we do not rank on it.
Agent evaluation on research-grade US tax questions using the Tax Agent Bench harness.
Evaluation results published by Vals AI. Published leaderboard aggregates and evaluation settings only. Dataset contents are not included.
Citation-adjusted all-pass accuracy on Tax Agent Bench — November 11, 2026
We mirror the published citation-adjusted all-pass accuracy view for Tax Agent Bench. Gemini 4 Hydrogen leads the public snapshot at 52.30%, followed by Gemini 4 Carbon (51.78%) and Claude Fable 5.1 (50.75%). We do not use these results to rank models overall.
Gemini 4 Hydrogen
xhigh reasoning
Gemini 4 Carbon
high reasoning
Claude Fable 5.1
Anthropic
85 modelsAgenticCurrentDisplay onlyUpdated November 11, 2026
Citation-adjusted all-pass accuracy table (85 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Gemini 4 HydrogenGooglexhigh reasoning | 52.30% | Not reported | Closed |
| 2 | Gemini 4 CarbonGooglehigh reasoning | 51.78% | Not reported | Closed |
| 3 | Claude Fable 5.1Anthropic | 50.75% | Not reported | Closed |
| 4 | Qwen3.8 Max InternalAlibaba | 49.66% | Not reported | Unknown |
| 5 | Gemini 4 ArgonGooglehigh reasoning | 49.22% | Not reported | Closed |
| 6 | Claude Opus 5Anthropic | 47.10% | Not reported | Closed |
| 7 | Claude Opus 5.5Anthropic | 44.54% | Not reported | Closed |
| 8 | DeepSeek V4.1 Flash 1017DeepSeekmax reasoning | 44.35% | Not reported | Open |
| 9 | GLM-5.3Z.AImax reasoning | 43.33% | Not reported | Open |
| 10 | Muse Wonder FullMetamax reasoning | 43.00% | Not reported | Closed |
| 11 | Muse Spark 1.3 MaxMetamax reasoning | 42.44% | Not reported | Unknown |
| 12 | Claude Sonnet 5.5Anthropic | 41.96% | Not reported | Closed |
| 13 | Inkling 2 SmallThinking Machines Lab0.99 reasoning | 38.91% | Not reported | Closed |
| 14 | Claude Fable 5Anthropic | 38.28% | Not reported | Closed |
| 15 | Muse EmberMetamax reasoning | 38.23% | Not reported | Closed |
| 16 | Muse Spark 1.3Metaxhigh reasoning | 37.89% | Not reported | Closed |
| 17 | Muse WonderMetamax reasoning | 36.27% | Not reported | Closed |
| 18 | Grok 4.7xAIxhigh reasoning | 35.75% | Not reported | Closed |
| 19 | Hy4Tencent | 34.40% | Not reported | Pending |
| 20 | Grok 4.6xAIhigh reasoning | 33.66% | Not reported | Closed |
| 21 | Kimi K3Moonshot AImax reasoning | 33.37% | Not reported | Open |
| 22 | Muse Spark 1.2Metaxhigh reasoning | 32.80% | Not reported | Closed |
| 23 | Hy4 previewTencent | 32.62% | Not reported | Open |
| 24 | MiMo-V2.6-ProXiaomi | 32.39% | Not reported | Open |
| 25 | MiniMax M3.1 FlashMiniMax | 32.14% | Not reported | Closed |
| 26 | Gemini 3.8 FlashGooglehigh reasoning | 32.12% | Not reported | Closed |
| 27 | Qwen3.8 MaxAlibaba | 32.09% | Not reported | Open |
| 28 | Claude Fable 5.2Anthropic | 32.09% | Not reported | Closed |
| 29 | Claude Haiku 5.5Anthropic | 31.92% | Not reported | Closed |
| 30 | GLM-5.3-FlashZ.AImax reasoning | 31.55% | Not reported | Open |
| 31 | GPT-5.6 SolOpenAImax reasoning | 31.09% | Not reported | Closed |
| 32 | Ember-1Fireworks | 31.08% | Not reported | Closed |
| 33 | Qwen3.8-27BAlibabaxhigh reasoning | 30.89% | Not reported | Open |
| 34 | Mistral Large 4 OssMistral AIhigh reasoning | 30.67% | Not reported | Unknown |
| 35 | Mistral Large 4Mistralhigh reasoning | 30.39% | Not reported | Open |
| 36 | GPT-5.6 TerraOpenAImax reasoning | 30.04% | Not reported | Closed |
| 37 | Claude Opus 4.8Anthropic | 30.00% | Not reported | Closed |
| 38 | Claude Sonnet 5Anthropic | 29.37% | Not reported | Closed |
| 39 | MiMo-V2.6-FlashXiaomi | 28.51% | Not reported | Open |
| 40 | DeepSeek V4 Flash 0731DeepSeekhigh reasoning | 28.07% | Not reported | Open |
| 41 | Step 5 PreviewStepFun | 27.27% | Not reported | Open |
| 42 | DeepSeek V4 Pro 0813DeepSeekmax reasoning | 26.76% | Not reported | Open |
| 43 | DeepSeek V4.1 Flash 1017DeepSeekmax reasoning | 26.63% | Not reported | Open |
| 44 | Muse Spark OpenMetaxhigh reasoning | 26.32% | Not reported | Open |
| 45 | Claude Sonnet 4.6Anthropic | 25.84% | Not reported | Closed |
| 46 | DeepSeek V4 Pro 0813DeepSeekmax reasoning | 25.17% | Not reported | Open |
| 47 | Kimi K3.1 ProMoonshot AImax reasoning | 24.87% | Not reported | Open |
| 48 | Grok 4.5xAIhigh reasoning | 24.86% | Not reported | Closed |
| 49 | GPT-5.6 LunaOpenAImax reasoning | 24.82% | Not reported | Closed |
| 50 | Claude Opus 4.7Anthropic | 23.74% | Not reported | Closed |
| 51 | Gemini 3.7 FlashGooglehigh reasoning | 22.80% | Not reported | Closed |
| 52 | GPT-5.5OpenAIxhigh reasoning | 21.76% | Not reported | Closed |
| 53 | GPT-6.1 SolOpenAImax reasoning | 21.75% | Not reported | Closed |
| 54 | Gemini 3.5 FlashGooglehigh reasoning | 21.65% | Not reported | Closed |
| 55 | GPT-6 AstraOpenAImax reasoning | 20.72% | Not reported | Closed |
| 56 | Ling 3.0 Flash Af Rc3Ant | 20.60% | Not reported | Unknown |
| 57 | MiniMax M3MiniMax | 20.59% | Not reported | Open |
| 58 | GPT-6 LunaOpenAImax reasoning | 19.95% | Not reported | Closed |
| 59 | Mercury 3Inceptionhigh reasoning | 18.17% | Not reported | Closed |
| 60 | Gemini 3.6 FlashGooglehigh reasoning | 18.04% | Not reported | Closed |
| 61 | GPT-6.1 TerraOpenAImax reasoning | 17.08% | Not reported | Closed |
| 62 | Qwen3.7 MaxAlibaba | 15.83% | Not reported | Closed |
| 63 | Inkling-SmallThinking Machines Lab0.99 reasoning | 15.39% | Not reported | Open |
| 64 | InklingThinking Machines Lab0.99 reasoning | 15.29% | Not reported | Open |
| 65 | GPT-6 SolOpenAImax reasoning | 15.00% | Not reported | Closed |
| 66 | Kimi K2.6Moonshot AI | 12.40% | Not reported | Open |
| 67 | Gemini 3 Flash PreviewGooglehigh reasoning | 11.84% | Not reported | Unknown |
| 68 | Nemotron 3.5 Super VL 120B A12BNVIDIAhigh reasoning | 11.55% | Not reported | Open |
| 69 | MiMo-V2.5-ProXiaomi | 11.50% | Not reported | Closed |
| 70 | Qwen3.7 PlusAlibaba | 10.47% | Not reported | Closed |
| 71 | Gemini 3.1 Pro PreviewGooglehigh reasoning | 9.29% | Not reported | Unknown |
| 72 | Beam 1.0Reflection | 8.81% | Not reported | Unknown |
| 73 | GPT-5.4 miniOpenAIxhigh reasoning | 8.81% | Not reported | Closed |
| 74 | MiMo-V2.5Xiaomi | 7.77% | Not reported | Closed |
| 75 | Qwen3.6 PlusAlibaba | 7.68% | Not reported | Closed |
| 76 | Gemini 3.5 Flash-LiteGooglehigh reasoning | 7.25% | Not reported | Closed |
| 77 | Grok 4.20 0309 ReasoningxAI | 6.95% | Not reported | Unknown |
| 78 | MiniMax M2.7MiniMax | 6.87% | Not reported | Open |
| 79 | Grok 4.3xAIhigh reasoning | 6.18% | Not reported | Closed |
| 80 | Claude Haiku 4.5 ThinkingAnthropic | 5.86% | Not reported | Closed |
| 81 | GPT-5.4 nanoOpenAIhigh reasoning | 4.97% | Not reported | Closed |
| 82 | Mistral Medium 3.5Mistral AIhigh reasoning | 4.48% | Not reported | Unknown |
| 83 | Gemini 3.1 Flash Lite PreviewGooglehigh reasoning | 3.11% | Not reported | Unknown |
| 84 | Nemotron Lightning 3p5 30b A3bFireworks AI | 2.59% | Not reported | Unknown |
| 85 | Mercury 2.5Inceptionhigh reasoning | 2.49% | Not reported | Closed |
How Tax Agent Bench is shown here
BenchLM mirrors the public Vals AI Tax Agent Bench leaderboard captured from https://www.vals.ai/benchmarks/tax_agent_bench and updated by Vals on November 11, 2026. The snapshot preserves overall scores, uncertainty, latency, cost-per-test metadata, and task-level scores where Vals publishes them.
Tax Agent Bench is display only on BenchLM. Vals proprietary or Vals-hosted aggregate views are useful context, but BenchLM does not use them as weighted ranking inputs or as a replacement for benchmark-native source records.
Snapshot
The published Tax Agent Bench snapshot places Gemini 4 Hydrogen first at 52.30%. The third row is 1.56 points behind. The broader top-10 range is 9.31 points, so many of the published results sit in a relatively narrow band.
85 models have been evaluated on Tax Agent Bench. The benchmark falls in the Agentic category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. Tax Agent Bench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About Tax Agent Bench
Year
2026
Tasks
193 private US tax test questions across six categories
Format
Citation-adjusted all-pass accuracy; Weighted Pass separate
Difficulty
Research-grade US tax questions
Primary accuracy is all-pass multiplied by a citation-quality factor of 0.7 to 1. An answer that misses any required check earns zero. Weighted Pass is a separate, must-pass-gated partial-credit metric. Both metrics remain display only.
Freshness and provenance
Version
Tax Agent Bench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Tax Agent Bench measure?
Agent evaluation on research-grade US tax questions using the Tax Agent Bench harness.
Which model leads the published Tax Agent Bench snapshot?
Gemini 4 Hydrogen currently leads the published Tax Agent Bench snapshot with 52.30% citation-adjusted all-pass accuracy. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on Tax Agent Bench?
The November 11, 2026 snapshot contains 85 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.