RULES WIKI

AI coding quota rules wiki

Every rule cites a verifiable source. Where an official number does not exist, we say so instead of inventing one.

How the rolling window is counted

Limits refill continuously from the moment each request is spent, not at a fixed midnight reset.

Claude Code and claude.ai run on a rolling window: capacity frees up progressively as the oldest usage ages out. Anthropic publishes a fixed window length (five hours on current plans) and shows you the exact refill timestamp in-product.
Codex usage is also consumed against a rolling allowance, layered with plan-level caps. Because OpenAI does not publish one universal number, the in-product meter beats any third-party estimate.
Practical consequence: a single large request is cheaper against the window than many small ones when the allowance counts messages.

Reasoning tokens count as output. Cache hits save latency and cost, but they are still requests.

Reasoning (thinking) tokens are billed as output tokens, so they consume the same allowance as a visible answer. Long reasoning traces are the most common reason a limit disappears faster than expected.
Prompt caching reduces latency and per-token cost on cache hits, but cached input is still input that passes through the endpoint and still counts against request and token rate limits.
If you need to stretch a window, lower the reasoning effort or shorten the context before you lower the model quality.

An on-demand refill credited to your account; you decide when to spend it. Expiry rules follow the official wording, not community lore.

A banked reset sits in your account until you trigger it, which makes it the fastest legal escape when you are blocked mid-task.
This tracker records every announced drop with its source link, so you can verify which events were global resets and which were compensation cards.
Because expiry terms can change, treat any fixed number of days you read online as unverified until you see it in your own dashboard.

Periodic drop events are a human decision on the provider side, historically clustered around incidents and launches.

Global resets are triggered by the provider, not by a schedule you can compute. Any site claiming a guaranteed time is guessing.
In the public record archived here, drops cluster around service incidents and major model rollouts, which is why this tracker feeds live incident data into its forecast.
The honest posture: treat the percentage as weather, not a timetable, and keep a fallback workflow ready.

LIVE FORECAST (SAME SOURCE AS HOME)

Built from the median drop interval, cooldown decay and live OpenAI incidents

Reset likelihood

72%

Countdown

-0d 2h 20m

Median cadence

3.0d