Arena finds top Anthropic and OpenAI coding models draw fewer user complaints

Arena analyzed 20,869 Agent Arena traces and reports that the five top coding models from Anthropic and OpenAI draw fewer user complaints than other agents.

Among sessions with complaints, failed repair loops come up 6.9 percentage points less often, design complaints fall 6.4 points, and failure to produce working results falls 5.8 points.

Arena also reports smaller drops in other complaint types: instructions followed less well (5.6 points less often), unfinished work (4.9 points), and breaking existing work (4.5 points).