APEX-Agents
Test agents on extended banking, consulting and legal work spread across files, communications and applications.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Mercor created the tasks with experienced professionals. Box helped design the file systems; Harvey advised on the legal worlds. The leaderboard describes 240 tasks across 31 workspaces.
Methodology
Version 1.1, announced September 8, 2026, revises task specifications, tools and grading to penalize answers that hedge among incompatible alternatives. Expert-written rubrics distinguish partial completion from fully correct work. Keep original and 1.1 results separate. The launch describes an open benchmark, while the current board also describes a private full set and open subset; their exact relationship is unspecified.