What will Codex actually cost, and what should you run?

CodexHow exists to answer one question honestly: what will running OpenAI Codex actually cost you, and what should you actually run? Not "roughly," not "it depends" — computed, from OpenAI's own published pricing and model pages, re-checked on a standing cadence rather than written once and left to rot.

Why this site exists

The Codex CLI ships with a default model that most people never think to question, a long-context repricing rule that reprices an entire request rather than the tokens over the line, and a rate-limit table that genuinely differs model to model in ways that matter for planning a CI pipeline or an agentic coding session. None of that is secret — it is published, in public, on OpenAI's own pages. It is just scattered across a pricing table, a dozen per-model pages, a CLI reference and a security guide, and none of those pages puts the pieces next to each other.

That is the entire job of this site: take numbers that are already public, put them in the same place, and compute the thing you actually wanted to know — what a described workload costs, whether a long session is about to cross an expensive threshold, and which model actually fits a task, not which one has the biggest number on its spec page.

Nothing here restates OpenAI's documentation. Every page on this site either computes an answer from a real workload you describe, measures two options against each other with the assumptions shown, or catalogues a specific mistake and its one-line fix. If a page could be replaced by copying a paragraph out of the official docs, it does not belong here.

What accuracy means on this site

Every figure that appears anywhere on CodexHow comes from a single facts module, keyed to a dated, cited OpenAI source. When a page states a number, it also states when that number was last checked and against which page — and the whole site fails its own build if any figure goes more than a handful of weeks without a re-check. Where OpenAI has not published something — a price, a rate limit, a repricing rule for a specific model — this site says so, in those words, rather than estimating a plausible-looking substitute. An admitted gap is more useful than a confident guess, because you can act on an admitted gap and you cannot safely act on a guess that looks like a fact.

This site is independent. It is not affiliated with, endorsed by, or sponsored by OpenAI, and it carries no OpenAI logo or wordmark anywhere. "Codex" and "ChatGPT" are used here only to describe what the site is about.

What you will actually find here

A set of calculators do the arithmetic you would otherwise do in a spreadsheet: pricing a single request, turning a Codex task into ChatGPT credits at any speed, comparing a flat subscription against metered API spend, checking a workload against the real published rate-limit tables, and — the tool this site was built around — pricing the same described workload on OpenAI's numbers next to your own Claude rates, with nothing declared a universal winner.

Underneath the tools sits a reference layer generated directly from the facts module: every priced model's context window, cache minimums and rate limits, from the GPT-6 generation back through GPT-5.4; the service tiers side by side, Ultrafast included; what each ChatGPT plan gets in Codex, and the credit rates behind it; which models are leaving Codex and what replaces them; and the sandbox and approval modes the CLI actually ships with. And underneath that, a growing catalogue of the specific mistakes developers hit — the CLI default nobody expects, the cache-write charge nobody budgets for, the approval flag that quietly changed — each with the cause and the one-line fix, not a general essay.

Frequently asked

Is this an official OpenAI site? No. CodexHow is independently run and clearly says so in the footer of every page and in full on the About page. Nothing here implies OpenAI wrote it or endorses it.

Does this site show ChatGPT plan prices? Yes — OpenAI's US list prices, with the date they were last checked, on Codex plans, prices and usage limits. They're a reference point, not a promise of what you pay: ChatGPT's own pricing page varies by region and currency, and Business is cheaper when billed annually. That's why every calculator here still starts empty and asks for your own figure, with the list prices one tap away.

Which model does the Codex CLI use by default? GPT-6.1 Sol, as of the date in the footer — it replaced GPT-5.6 Sol this autumn, and it isn't the model named for Codex. Run /status in a session to see what you're actually running.

What changed recently? A lot: a new model generation, a new CLI default, an Ultrafast tier, a reshaped set of Pro plans and several retirements. The running table of pricing changes lists them in order, along with what this site itself got wrong.

How often are the numbers checked? On a standing weekly cadence, and mechanically enforced: the site refuses to build if any figure has gone unchecked for too long. The exact date is on every page that states a number.

What happens when OpenAI doesn't publish something? The page says "not published," names which OpenAI pages were actually checked, and stops there. It never estimates a figure to fill the gap.