Codex usage limits work on rolling 5-hour windows, shared across every Codex surface (CLI, IDE, web, iOS) on one plan allowance. On ChatGPT Plus ($20/mo) OpenAI's own estimates run from 3–30 messages per 5 hours on the top model (GPT-6 Astra) up to 250–2,000 on the lightest (GPT-5.6 Luna); Pro ($100–200/mo) is 5x or 20x that. When you see 'You've hit your usage limit. Try again later.', your options are: wait for the window, switch to a lighter model, buy credits, upgrade, or move routine work to another provider.
| Plan | Price | Codex allowance |
|---|---|---|
| Free | $0 | Limited Codex access |
| Go | $8/mo | Entry allowance |
| Plus | $20/mo | Standard allowance — the community default |
| Pro | $100–200/mo | 5x or 20x Plus |
| Business | $20/user/mo (annual) | Plus-level per seat |
From OpenAI's official pricing docs, 2026-09-03. API pay-as-you-go has no windows — the meter is the limit.
The limits are estimates, not contracts — OpenAI publishes ranges, tightened them in 2026, restored them after backlash, and can change them again. If your workflow depends on a guaranteed volume, a windowed subscription cannot promise it; only a metered API or a budgeted flat plan can.
Codex CLI supports custom OpenAI-compatible providers, and about a quarter of Standard Compute's active customers run it against our API: base URL api.stdcmpt.com/v1, one flat plan from $19/mo, smart routing, no 5-hour windows. The pattern that works: keep your ChatGPT plan for what needs OpenAI's top models, run the volume work through a flat budget.
There is no fixed reset time — windows roll continuously over 5 hours, so capacity frees up as older usage ages out. The usage dashboard shows where you stand. Pro plans also reference weekly caps for sustained heavy use.
Almost always the model: top-tier models like GPT-6 Astra allow as few as 3–30 messages per 5-hour window on Plus. Agent runs also consume multiple requests per visible message — one 'message' of agentic work can be dozens of calls.
The 20x tier only pays off for parallel, sustained agentic work — the same logic as Claude Max 20x. Most heavy users land on the $100 5x tier, and community experience says even that mostly matters on the top models.