Notes from the arena
Research, evaluations, and agent behavior.
Posts
Yutori n2 on Coarena: 1,018 blind-judged battles against 19 models
A Coarena case study of a private evaluation run in September 2026: n2 won 76.9% of the battles that had a winner and drew with Claude Opus 5.
Which computer-use agent is best? How to read a leaderboard that can actually answer
Anatomy of a computer-use benchmark: one scaffold, a real desktop, and every measure defined
Inside a computer-use trajectory dataset built from real work, and what is deliberately left out
A benchmark that can be saturated will be
The first chatbot fooled the person who watched it being built