APEX-Agents leaderboard

What APEX-Agents measures

APEX-Agents was created by Mercor to test whether AI agents can do long-horizon, cross-application work in professional services. It contains 480 tasks, 160 each in investment banking, management consulting and corporate law, set in 33 simulated work environments that each average more than 160 files. Industry professionals with five to ten years of experience at top-tier firms built the worlds and tasks, each around a realistic scenario such as a week-long consulting engagement.

Agents use the tools a human professional would: documents, spreadsheets, presentations, PDFs, email, chat, a calendar, a file system and code execution, with web search disabled so runs are reproducible. Each task has an expert-written rubric of one to ten binary criteria, and a judge model checks each criterion against the agent's output. A task passes only if every criterion is met, and the headline number is pass@1. Mercor estimates the tasks take a human professional about 1.8 hours on average.

How to read this page

This is the closest benchmark on the site to "can an agent do a white-collar job step". The cost chart uses output price per 1M tokens, but an agent run here also reads large file sets and makes many tool calls, so input tokens and caching matter more than on a one-shot question. Use the price axis to see which tier of model is good enough, and then price a realistic trajectory with the agent cost calculator rather than trusting the per-token figure alone.

Limitations

Passing requires every rubric criterion, so the metric is strict and a single missed detail fails a task. Grading uses a judge model, which adds its own bias and noise. The tasks are confined to three professions and to simulated environments, so results may not transfer to other kinds of work or to messier real workplaces. The benchmark is open source (tasks and worlds under CC-BY, with the Archipelago environment released openly), which helps reproducibility but also means the tasks can leak into training data.