Atlas

Benchmarks

← All benchmarks

Tax Agent Bench — Citation-Adjusted Forms & Filings All-Pass

Professional Work · 2026-10-08

Citation-adjusted all-pass credit on 30 held-out forms-and-filings questions, under the October 8 headline scoring protocol.

Top models (higher is better)

ModelScore
DeepSeek V4.1 Flash 101733.3
Claude Opus 533.2
Claude Fable 5.130.0
Grok 4.730.0
Muse Wonder Full30.0
Claude Fable 526.7
Gemini 4 Argon26.7
Hy4 Preview26.7
MiniMax M3.1 Flash26.7
Mistral Large 4 Preview26.7
Muse Spark 1.225.3
Muse Spark 1.324.7
GLM-5.3 Flash23.3
Grok 4.623.3
Muse Wonder23.3
Claude Haiku 5.520.0
Ember-120.0
Gemini 3.8 Flash20.0
GLM-5.320.0
GPT-5.6 Sol20.0
Kimi K320.0
MiMo-V2.6-Pro19.9
Sonnet 519.9
MiniMax M319.9
MiMo-V2.6-Flash17.0
Gemini 3.5 Flash16.7
GPT-5.6 Terra16.7
Grok 4.516.7
Qwen3.8 27B16.7
DeepSeek-V4-Flash-073115.7
Step 5 Preview14.7
Opus 4.813.3
GPT-5.6 Luna13.3
GPT-6 Luna12.8
DeepSeek V4.1 Flash12.3
DeepSeek V4 Pro 081312.3
Mercury 312.3
Muse Spark Open11.7
Inkling-Small10.3
Gemini 3.7 Flash10.0
Loading Atlas data…