Best coding agent fully finishes only 26.4% of real user requests

The tasks come from 37,463 requests across 26,314 sessions, picked where users had to rephrase a request or roll back a change. A small job might mean reading a figure off a receipt and fixing every total built on it, and a big one means reworking a whole project. Grading uses two AI judges: one works the app in a browser, the other checks each pass for mock-ups or answers typed in rather than worked out.

Claude Opus 5 ranks second and Qwen3.8-Max third, while Grok 4.6 finishes tenth of ten. Every model does worse when a task mixes three or four kinds of change, such as layout, design, interaction and content. GLM-5.2 used nearly six times GPT-6 Astra’s tokens (the text units models bill by) on small changes and still scored lower.

Our price sheet, checked 8 November, shows Claude Pro at $20 a month with Claude Code, Anthropic’s coding agent, included.

For a marketer who builds landing pages with an agent, the 26.4% figure means a human check on every change. The agent usually gets most of a revision right and misses at least one detail, such as a total that no longer adds up.