Usage limits
Every coding-agent subscription is sold on a session window and a weekly cap, and none of the providers publishes what those hold in tasks. These pages explain each harness's limits from the provider's own pages and put a measured number on them from our weekly runs.
Provider pages reviewed Sep 14, 2026; figures from the W38 2026 grading run.
Claude Code usage limits
Anthropic · measured on Claude Max 20x
about 14 runs a weekClaude Fable 5.1 on Claude Max 20x, about 270 tasks
Two windows on one meter. When the limits reset. What Anthropic publishes and what it does not.
Codex CLI usage limits
OpenAI · measured on ChatGPT Pro 5x
about 33 runs a weekGPT-5.6 Sol on ChatGPT Pro 5x, about 630 tasks
A 5-hour window with a weekly cap behind it. When the limits reset. What OpenAI publishes and what it does not.
Antigravity CLI usage limits
Google · measured on Google AI Pro
about 5 runs a weekGemini 3.8 Flash on Google AI Pro, about 94 tasks
A 5-hour refresh under a weekly limit. When the limits reset. What Google publishes and what it does not.
One run is one pass of our fixed task set (real-world issues plus spec planning, 19 tasks) on the plan our subjects grade on. Each harness page tabulates every plan of that provider and every graded model, with sibling plans estimated from the provider's advertised multiples and marked est.