Claude Opus 5 passes 23.9% of coding-agent benchmark evaluations, DAIR.AI says

DAIR.AI reports that Claude Opus 5 running under Claude Code passes 23.9% of evaluations, compared with 82.2% for an expert human reference. It describes tasks that give a developer agent business records, a client who can answer questions, a production API and an inherited codebase, with hard limits on serving cost and model choice. The resulting customer service agents are scored by deploying them against held-out simulated users. DAIR.AI identifies shallow searches of business records, little client communication and limited experimentation with architecture or serving spend as failures—agents ship the first design that runs.