Tax Agent Bench (Vals AI) leaderboard
- Claude Fable 5.1 leads
- 49.2%
- models measured
- 68
- lead over #2
- +3.6 pts
- median
- 25.5%
Leader on Tax Agent Bench (Vals AI) · read October 21
On October 21, 2026, 68 of our catalog's models had a Tax Agent Bench (Vals AI) result, and Anthropic Claude Fable 5.1 led with 49.2%.
US corporate tax research by an agent, graded on the answer and its citations. The score as Vals AI publishes it.
Role in the index
We collect this benchmark and show it as supporting evidence. It is not part of the Vibe Coding Index, because it measures something other than coding and the index is about coding. Read the methodology
| # | Model | Result | Cost per test | Duration, as printed | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.1Anthropic | 49.2% | $13.18 | 28m 34s | 89.1 | $50.00 | |
| 2 | Claude Opus 5.5Anthropic | 45.6% | $15.09 | 1h 07m | 92.2 | $20.00 | |
| 3 | Claude Opus 5Anthropic | 45.5% | $5.14 | 16m 48s | 81.8 | $25.00 | |
| 4 | Gemini 4 ArgonGoogle | 44.6% | $3.47 | 16m 26s | 80.5 | $10.00 | |
| 5 | DeepSeek V4.1 Flash 1017DeepSeek | 44.4% | $0.46 | 21m 04s | 83.3 | $0.60 | |
| 6 | Claude Sonnet 5.5Anthropic | 42.0% | $9.25 | 59m 21s | 88.9 | $10.00 | |
| 7 | GLM 5.3Z.ai | 39.7% | $1.80 | 50m 53s | 73.1 | $1.40 | |
| 8 | Claude Fable 5Anthropic | 38.3% | $6.70 | 14m 27s | 82.1 | $50.00 | |
| 9 | Muse WonderMeta | 36.3% | $1.88 | 8m 13s | 64.5 | $20.00 | |
| 10 | Muse Spark 1.3Meta | 36.0% | $0.32 | 8m 25s | 79.7 | $4.25 | |
| 11 | Grok 4.7xAI | 34.2% | $1.80 | 16m 36s | 75.0 | $6.00 | |
| 12 | Kimi K3Moonshot AI | 33.4% | $2.79 | 40m 15s | 61.8 | $11.936 | |
| 13 | Grok 4.6xAI | 33.1% | $0.98 | 10m 58s | 76.4 | $6.00 | |
| 14 | Muse Spark 1.2Meta | 32.8% | $0.27 | 1m 60s | 65.1 | $4.25 | |
| 15 | Hy4 previewTencent | 32.6% | $0.59 | 41m 04s | 61.6 | $2.501 | |
| 16 | MiMo-V2.6-ProXiaomi | 32.4% | $0.13 | 25m 48s | 72.6 | $0.87 | |
| 17 | Gemini 3.8 FlashGoogle | 32.1% | $0.86 | 2m 58s | 70.5 | $4.125 | |
| 18 | Qwen3.8 Max (0902)Qwen | 32.1% | $1.55 | 40m 36s | 63.5 | $6.00 | |
| 19 | Claude Haiku 5.5Anthropic | 31.9% | $1.16 | 49m 34s | 75.9 | $0.50 | |
| 20 | GLM 5.3 FlashZ.ai | 31.6% | $0.14 | 21m 37s | 71.0 | $0.50 | |
| 21 | GPT-5.6 SolOpenAI | 31.1% | $7.51 | 44m 41s | 76.7 | $20.00 | |
| 22 | Ember-1Fireworks | 31.1% | $2.97 | 30m 46s | Not benchmarked | $15.00 | |
| 23 | Qwen3.8 27BQwen | 30.9% | $0.82 | 22m 14s | 73.6 | $1.35 | |
| 24 | Mistral Large 4Mistral AI | 30.4% | $2.72 | 57m 18s | 55.3 | $4.18 | |
| 25 | GPT-5.6 TerraOpenAI | 30.0% | $4.83 | 1h 01m | 69.0 | $12.00 | |
| 26 | Claude Opus 4.8Anthropic | 30.0% | $1.69 | 7m 59s | 66.0 | $25.00 | |
| 27 | Claude Sonnet 5Anthropic | 29.4% | $1.79 | 19m 19s | 63.5 | $10.00 | |
| 28 | MiMo-V2.6-FlashXiaomi | 28.5% | $0.05 | 17m 19s | 61.9 | $0.28 | |
| 29 | DeepSeek V4 Flash 0731DeepSeek | 28.1% | $0.16 | 5m 37s | Not benchmarked | $1.32 | |
| 30 | Step 5 PreviewStepFun | 27.3% | $0.55 | 21m 29s | 71.4 | $2.70 | |
| 31 | DeepSeek V4 Pro 0813DeepSeek | 26.8% | $0.73 | 26m 37s | 54.6 | $1.98 | |
| 32 | DeepSeek V4.1 FlashDeepSeek | 26.6% | $0.13 | 5m 49s | 77.2 | $0.525 | |
| 33 | Muse Spark OpenMeta | 26.3% | $0.88 | 11m 42s | 62.1 | $4.25 | |
| 34 | Claude Sonnet 4.6Anthropic | 25.8% | $1.19 | 11m 47s | 57.8 | $15.00 | |
| 35 | DeepSeek V4 Pro 0423DeepSeek | 25.2% | $0.72 | 20m 45s | 35.3 | $1.9105 | |
| 36 | Grok 4.5xAI | 24.9% | $0.47 | 4m 44s | 75.5 | $6.00 | |
| 37 | GPT-5.6 LunaOpenAI | 24.8% | $0.47 | 42m 46s | 61.5 | $1.20 | |
| 38 | Claude Opus 4.7Anthropic | 23.7% | $1.27 | 4m 39s | 51.1 | $25.00 | |
| 39 | Gemini 3.7 FlashGoogle | 22.8% | $0.48 | 1m 27s | 69.9 | $3.75 | |
| 40 | GPT-5.5OpenAI | 21.8% | $4.07 | 15m 52s | 67.8 | $30.00 | |
| 41 | GPT-6.1 SolOpenAI | 21.8% | $1.32 | 20m 14s | 88.7 | $10.00 | |
| 42 | Gemini 3.5 FlashGoogle | 21.6% | $0.86 | 3m 53s | 42.8 | $9.00 | |
| 43 | GPT-6 AstraOpenAI | 20.7% | $5.80 | 15m 01s | 88.7 | $50.00 | |
| 44 | MiniMax M3MiniMax | 20.6% | $0.16 | 5m 15s | 57.8 | $1.20 | |
| 45 | GPT-6 LunaOpenAI | 19.9% | $0.19 | 33m 16s | 71.4 | $0.50 | |
| 46 | Gemini 3.6 FlashGoogle | 18.0% | $0.36 | 1m 45s | 48.3 | $3.75 | |
| 47 | Qwen3.7 MaxQwen | 15.8% | $0.44 | 3m 30s | 45.1 | $4.425 | |
| 48 | T Inkling SmallThinking Machines | 15.4% | $0.09 | 6m 25s | 29.9 | $1.20 | |
| 49 | T InklingThinking Machines | 15.3% | $0.35 | 13m 31s | 49.7 | $4.05 | |
| 50 | GPT-6 SolOpenAI | 15.0% | $2.28 | 17m 32s | 82.8 | $10.00 | |
| 51 | Kimi K2.6Moonshot AI | 12.4% | $0.57 | 15m 24s | 44.2 | $3.41 | |
| 52 | Gemini 3 Flash PreviewGoogle | 11.8% | $0.10 | 1m 41s | 75.8 | $3.00 | |
| 53 | MiMo-V2.5-ProXiaomi | 11.5% | $0.04 | 6m 17s | 40.0 | $0.609 | |
| 54 | Qwen3.7 PlusQwen | 10.5% | $0.09 | 5m 10s | Not benchmarked | $1.28 | |
| 55 | Gemini 3.1 Pro PreviewGoogle | 9.3% | $0.30 | 1m 41s | 35.3 | $12.00 | |
| 56 | GPT-5.4 MiniOpenAI | 8.8% | $0.81 | 16m 46s | 28.2 | $4.50 | |
| 57 | MiMo-V2.5Xiaomi | 7.8% | $0.02 | 8m 21s | 34.0 | $0.238 | |
| 58 | Qwen3.6 PlusQwen | 7.7% | $0.30 | 6m 18s | 37.4 | $1.95 | |
| 59 | Gemini 3.5 Flash LiteGoogle | 7.2% | $0.08 | 47 s | 33.6 | $2.50 | |
| 60 | Grok 4.20xAI | 7.0% | $0.12 | 1m 29s | Not benchmarked | $2.50 | |
| 61 | MiniMax M2.7MiniMax | 6.9% | $0.12 | 5m 03s | 28.4 | $1.20 | |
| 62 | Grok 4.3xAI | 6.2% | $0.12 | 1m 38s | 22.5 | $2.50 | |
| 63 | Claude Haiku 4.5Anthropic | 5.9% | $0.23 | 2m 35s | 49.8 | $5.00 | |
| 64 | GPT-5.4 NanoOpenAI | 5.0% | $0.07 | 3m 40s | Not benchmarked | $1.25 | |
| 65 | Mistral Medium 3.5Mistral AI | 4.5% | $1.44 | 7m 11s | 9.2 | $7.50 | |
| 66 | Gemini 3.1 Flash Lite PreviewGoogle | 3.1% | $0.01 | 19 s | 8.1 | $1.50 | |
| 67 | Nemotron 3.5 LightningNVIDIA | 2.6% | $0.03 | 19m 40s | Not benchmarked | $0.20 | |
| 68 | Mercury 2.5Inception | 2.5% | $0.02 | 30 s | Not benchmarked | $0.15 |
Only models in our catalog are listed, and each appears once with its best published run. 132 rows were read from the source on October 21, 2026, and rows that match no catalog model are left out. Cost per test is what Vals AI reports one test cost on the run behind the result, in US dollars. Every other column is a figure the board publishes for the same run.