chatgpt.com/codex/pricing, line by line, and the number it never shows you.

The full plan ladder

Here is every plan from the official pricing page, with the Codex allowance attached. The prices are flat monthly fees. The usage column is where the real spread lives.

Plans and prices verified against developers.openai.com/codex/pricing on 2026-05-15. The chatgpt.com/codex/pricing page sits behind a bot wall, so the developer subdomain is the readable mirror of the same ladder.

The pricing page quotes a range, not a number

This is the detail almost every pricing write-up flattens out. When the official page describes the Pro tiers, it does not give you a message count. It gives you a band. The Pro 5x tier is quoted as 80 to 400 local messages on GPT-5.5. The Pro 20x tier is quoted as 300 to 1600. Those are five-times wide ranges, and the page is explicit about why: usage varies by model and task complexity.

So the honest reading of “5x more usage” is not “a fixed bigger number”. It is “a wider band, and where you land inside it depends on what you actually do.” A handful of short prompts on a cheaper model puts you near 400. One long agentic refactor on GPT-5.5, with the model re-reading files and chaining tool calls, drops you toward 80. The pricing page cannot resolve that for you. It is a menu, not a meter.

That is the structural reason a flat monthly fee feels unpredictable. You paid one clean number. You then live inside a range that the number does not pin down, gated by a rolling 5-hour window and a weekly limit on top.

The April 2, 2026 change: from messages to tokens

On April 2, 2026, OpenAI moved Codex billing off per-message pricing and onto API-style token rates. Consumption is now measured as credits per million input tokens, cached input tokens, and output tokens. Credits stayed the unit you buy and spend, but what draws them down changed: it is now token volume, the same way an API bill works. The shift applied to new and existing Plus, Pro, and Business workspaces, with Enterprise migrating on its own schedule.

The pricing page publishes the credit rate for each model. The two most relevant ones for daily Codex work look like this:

Rates are credits per one million tokens, verified on developers.openai.com/codex/pricing on 2026-05-15. GPT-5.4 is roughly half the credit cost of GPT-5.5 per token, which is why model choice inside a session moves your effective spend as much as message count does.

Once your included plan usage is spent, further work draws down purchased credits at these rates. That is the mechanism that keeps a long session running past the plan limit instead of hard-stopping. The trade is that your monthly cost is no longer fixed. It is the flat fee plus however many credits the work above the line consumed, and the pricing page does not show you that running total.

Codex pricing and Claude pricing are now the same shape

If you are pricing out Codex, you are almost certainly weighing it against Claude Code, or already running both. Worth knowing: after the April change, the two pricing models are close to isomorphic. A flat plan, a rolling 5-hour window, a weekly limit, and a metered overage path. The prices line up too.

The interesting row is the last one. Both tools gate you the same way, but they put the live readout in different places. Codex prints a usage summary inside its own CLI when you run /status. Claude does not print anything comparable during an active agent loop. The only place to see your real Claude position is claude.ai/settings/usage, a web page you have to remember to open and refresh. So in a two-tool setup, the Codex side has a meter and the Claude side goes dark.

Why this matters for the Claude side of your bill

claude-meter exists for exactly that dark side. It is a free, open-source macOS menu bar app and browser extension that reads your Claude Pro or Max usage and shows it live. It does not track Codex, and it makes no attempt to. Codex already has /status. What was missing was the matching readout on the Anthropic side.

The browser extension makes one HTTPS request per minute to claude.ai using the session cookies your browser already holds, then pushes the snapshot to the menu bar app over a localhost bridge. No cookie paste, no telemetry, no separate login. The numbers come from the same usage endpoint that backs claude.ai/settings/usage, so they match the page exactly. There is also a CLI:

That is the readout the Codex pricing page has a sibling for and the Claude settings page only offers behind a manual refresh. Pair it with Codex’s /status and both of your flat-rate plans finally have a live meter, so you see a window filling up before it walls a task rather than after. If you run both agents in rotation, the token-juggling guide walks through the exact handoff that asymmetry makes possible.

The short version

Codex pricing is plan pricing. You pick a ChatGPT plan, $20 for Plus up to $200 for the top Pro tier, and Codex comes with it. The April 2 change turned the overage into token-metered credits. None of that is hard to read off the official page.

The part the page does not give you is your position. A 5x plan is a wider range, not a fixed number, and the 5-hour and weekly windows decide when a long task stops. That is true on the Codex side, where /status at least shows you, and on the Claude side, where nothing does until you install something that reads it. The sticker price is the easy half. The meter is the half worth watching.

Frequently asked