How much does one percent of Codex's 5-hour window cost?

There is no published answer: OpenAI does not state the 5-hour window in tokens or dollars, and calls its per-plan message counts estimates, not fixed limits. But Codex writes its own reading of the window, the percentage used, into the session files on your machine next to the tokens each turn used. So one percent is measurable on your account: the tokens spent between two readings of the same window, divided by how far the percentage moved. Priced at OpenAI's API list prices, that is a dollar figure, and a step in it means the window now gives you more or fewer tokens per point. npx seatledger rates --client codex tracks it day by day.

What OpenAI publishes (7 October 2026)

Plus and Standard Business have a five-hour period. OpenAI's estimates of local messages per five hours on Plus, by model:

Pro plans currently have no five-hour limit, and weekly limits may also apply. Fast mode uses included limits at 2.5× the Standard rate for the same model, and GPT-6 Astra Ultrafast at 8×. Inside a session, /status shows what is left. A range that wide is why a measured number is worth having: what one message costs depends on the model, the context, reasoning, tools and caching.

The method

Codex records rate-limit readings in its rollouts: for the primary (5-hour) and secondary (weekly) windows, the percentage used, the window length and when it resets, and the plan. seatledger groups the 5-hour readings by the window they belong to (its reset time), takes the first and last reading of each, adds up the tokens of this machine's Codex requests between them, and divides by the points the percentage moved. Days where the window moved less than two points are left out. Then it runs the same step detector it uses for Claude Code: a change of 30% or more that holds for days on both sides is reported with the date and the Codex version in use.

For a dollar figure, divide the API-equivalent dollars of npx seatledger report --client codex --by model by its tokens and multiply by the tokens per 1%. That is what one point of the window would cost at API list prices; it is not what your plan costs per point.

$ npx seatledger rates --client codex
  • From 22 Sep (Codex 0.159.0, the first day on it), tokens per 1% of the 5-hour window halved across Codex models — consistent with a smaller allowance per token: the same tokens now use more of the 5-hour window (usage on other devices on the same account also lowers this). It starts with Codex 0.159.0 (previously 0.158.0).

What it cannot see

The percentage is account-wide: Codex on another machine, cloud tasks and other features that share the allowance move it too, which makes this machine's tokens per 1% look smaller. On Pro there is no five-hour window to divide by; seatledger still shows the weekly reading Codex writes. A switch to Fast mode lowers tokens per 1% by the multiplier above without anything else changing.

What people reported

In openai/codex#28879 a Plus user showed, from their own session logs, the 5-hour percentage used per token rising about tenfold or more from around 16 June 2026 with smaller prompts and little reasoning. The issue was closed as completed on 23 August 2026. It was measured by hand from the same readings seatledger divides.

$ npx seatledger