About CodexHow
CodexHow is a small, independently run reference site for developers building with OpenAI Codex and the Codex CLI. It exists because the numbers that matter for planning a Codex workload — pricing, context windows, cache economics, rate limits, sandbox and approval behaviour — are all genuinely published by OpenAI, but scattered across a pricing table, a stack of per-model pages, a CLI reference and a security guide that never sit next to each other. This site puts them next to each other and computes the answer you actually wanted.
Independence
CodexHow is not affiliated with, endorsed by, or sponsored by OpenAI. "Codex," "ChatGPT" and "OpenAI" are used throughout this site only to describe what it is about — never to suggest this site speaks for OpenAI or that OpenAI has reviewed anything published here. No OpenAI logo, wordmark, or icon appears anywhere on this site, in any tool, table, or page.
Accuracy {#accuracy}
Every number this site publishes lives in one internal module, and every entry in it carries the date it was last checked and the exact OpenAI page it was checked against. Pages render numbers from that module — nothing is typed by hand into an article, a how-to guide, or a blog post. That single rule is what makes a site-wide re-check possible at all: change a number in one place, and every page that cites it updates together.
The site's own build process enforces a freshness limit. If any figure in the facts module has gone unchecked for too long, the build fails outright rather than deploying quietly stale numbers. That is a deliberate choice: a confidently out-of-date developer-tools site is worse than one that visibly refuses to ship until someone re-checks it.
Where OpenAI has not published a figure — a price for a specific model, a rate-limit table, a long-context repricing rule for a given model — this site states plainly that it is not published, and names which OpenAI pages were actually checked before making that claim. That last part matters: this site's own first research pass got that wrong three times in one afternoon, reporting a figure as unpublished after checking only one of several pages where OpenAI actually states it. Every "not published" claim on this site now records what was checked, specifically so that mistake is visible and correctable rather than quietly repeated.
Nothing on this site estimates a number OpenAI hasn't published. An admitted gap, clearly labelled, is more useful to a developer making a real budgeting decision than a plausible-looking guess.
No subscription prices
CodexHow does not publish a ChatGPT or Codex plan price anywhere, for the same reason it doesn't publish a Claude plan price: neither is available from a source this site treats as citable and authoritative, and both vary by region, currency and billing period in ways a single number can't represent honestly. Every tool that needs a subscription figure asks you for your own instead, so the result is correct for your actual plan rather than a generic average.
What this site will never do
It will not restate OpenAI's own documentation as if it were new content — every page here computes something, compares something with the assumptions shown, or catalogues a specific mistake with its fix. It will not declare a universal winner in any cross-provider comparison; it will state which side is cheaper for the specific workload you described, and by how much, and leave the general question open because there genuinely isn't one honest general answer. And it will not quietly go stale: if the weekly re-check ever stops happening, this site should come down rather than keep serving numbers nobody is checking anymore.
Who runs this
CodexHow is one part of a small portfolio of independently maintained, narrowly scoped reference sites, including its closest sibling, a similarly built site for the Claude API and Claude Code. Ordinary advertising may appear on this site to cover hosting; it never decides what a page says, and no page here has ever been shown to an advertiser before publication.