AA-Briefcase

Artificial Analysis's private 91-task agentic benchmark of long-horizon professional knowledge work, scored as combined Elo from rubric success, analytical quality, and presentation.

Also known as: AA Briefcase, Artificial Analysis Briefcase

active

This benchmark is in the default catalogue: its identity, protocol, current model coverage and dated results were verified by a reviewer who opened the sources.

Recorded reasons:

Categoryagentic
Subcategorylong-horizon professional knowledge work with file deliverables
Page statusactive
Metriccombined Elo, Index-normalized as clamp((Elo − 500) / 2000)
Directionhigher_is_better
Unit%
Dataset size91
PublisherArtificial Analysis

What it measures

AA-Briefcase measures whether a model can complete realistic, multi-week professional knowledge-work projects as file deliverables. The scored set is 91 tasks across four private scenarios (data science, product management, banking operations, and heavy industry strategy). Each scenario is a linked weekly workflow with thousands of source files. The agent must produce artefacts such as spreadsheets, presentations, memos, and PDFs in an offline sandbox, without live user feedback. Tasks currently run independently, so a model does not carry its own prior submissions into later weeks.

Task format

One independent Stirrup/E2B run per task (up to 500 turns). The agent receives the scenario and week overviews, the task brief, and mounted source files, then submits named deliverable files. No internet. Vision models also get a view-image tool. One repeat. Official scores use the private 91-task set, not the public Lite scenario.

Verified results

Each row was checked against its source by a reviewer.

ModelScoreEvidence dateSource kindLink
GPT-6 Astra (max)53.0 normalized Elo percent2026-09-04 publishedindependent_evaluatorsource
GLM-5.3 (max)51.0 normalized Elo percent2026-09-04 publishedindependent_evaluatorsource

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub