Tax Agent Bench
Results
26 of 41 tracked models have a result.
- 1
Anthropic · Max effort
77.6% - 2Gemini 4 ArgonLimited
Google DeepMind · High effort
76.2% - 3
Anthropic · Max effort
73.4% - 4GLM-5.3Open
Z.ai · Max effort
73.1% - 5
Meta · Max effort
72.4% - 6
Anthropic · Max effort
70.5% - 7Kimi K3Open
Moonshot AI · Max effort
68.7% - 8
Google DeepMind · High effort
66.8% - 9
Alibaba
66.0% - 10
SpaceXAI · xHigh effort
65.6% - 11
OpenAI · Max effort
65.2% - 12MiMo-V2.6-ProOpen
Xiaomi
64.9% - 13Hy4 previewOpen
Tencent
63.7% - 14
OpenAI · Max effort
63.3% - 15
Mistral AI · High effort
63.3% - 16
DeepSeek · High effort
62.5% - 17
OpenAI · Max effort
62.3% - 18
Anthropic · Max effort
62.2% - 19MiMo-V2.6-FlashOpen
Xiaomi
59.9% - 20
StepFun
59.5% - 21
OpenAI · Max effort
58.9% - 22DeepSeek V4 ProOpen
DeepSeek · Max effort
58.7% - 23MiniMax-M3Open
MiniMax
49.7% - 24InklingOpen
Thinking Machines Lab
43.5% - 25Inkling-SmallOpen
Thinking Machines Lab
41.5% - 26
Google DeepMind · High effort
37.8%
15 tracked models without a result
Command A+, Gemini 3.5 Flash-Lite, Gemma 4 31B, GLM-5.3-Flash, gpt-oss-120b, Hy3, Kimi K2.7 Code, MAI-Thinking-1, Mistral Medium 3.5, Mistral Small 4, Muse Glimmer 30B, Nemotron 3 Ultra, Nova 2 Lite, Qwen3.8-27B, Qwen3.8-Flash.
Missing means the evaluator has not published a run, not that the model would score poorly.