Inference
EN
Nous Research

Hermes Bench: 150 agent tasks graded on the files and state the agent leaves behind

Tests agents on 150 tasks across 25 categories, including Hermes skills, research, diagrams and art, memory, tool use and safety. Each agent starts in a workspace with real files, some tasks add follow-up turns, and the grader checks the files and state left behind.

ModelsModels with a result.16
Top resultBest result on this benchmark.—
Top-3 spreadPoints from first to third.—
YearYear of release.2026

Result and price

15 of 16 results

Results

16 results
#ModelResultThe score from the source.
—Claude Opus 5.576.7%
—Grok 4.774.4%
—GPT-6.1 Sol73.9%
—GPT-6 Astra73.7%
—GPT-6 Sol73.5%
—Claude Sonnet 5.573.5%
—GLM-5.371.8%
—Kimi K371.5%
—Claude Haiku 5.571.0%
—Qwen3.8 Max70.5%
—GPT-6 Luna70.5%
—Step 5 Preview69.2%
—GLM-5.3-Flash68.9%
—Gemini 3.8 Flash62.2%
—Hy4 preview58.6%
—Ling 3.0 Flash53.8%