Atlas

Benchmarks

← All benchmarks

MissionReport

Professional Work

micro1's MissionReport: 20 government status-reporting assignments (weekly activity, program status and after-action reports and performance assessments for Army, Air Force, Navy, Marine Corps and Department of Defense offices), each written by an expert with a role, a reader, the working files and the office's template. The agent must save a finished Word document from an OpenCode sandbox, and an agentic Claude Opus 5.5 judge decides each criterion of the expert's rubric. A report scores positive points earned minus penalty points over the task's positive points, bounded to 0-100%, and the score is the mean over tasks of three reports.

Top models (higher is better)

ModelScore
Gemini 4 Argon64.1
Claude Sonnet 5.563.8
Claude Opus 5.563.5
Claude Fable 5.163.3
Muse Wonder62.1
GPT-6 Astra61.8
GPT-6.1 Sol61.0
Grok 4.759.5
Gemini 3.8 Flash54.3
Loading Atlas data…