OpenAI Codex cloud pricing explained: a rate limit, not a rate
What Codex cloud costs on each ChatGPT plan, why OpenAI publishes a message range instead of a price, and what the API-key path gives up.
Codex cloud does not have a price. It has a plan. OpenAI bundles its cloud coding agent into ChatGPT subscriptions, so a task costs you nothing at the moment you run it and the bill never itemizes one. What you buy instead is throughput: a rate limit that scales with the tier, published as a range of messages per five hours rather than a dollar figure. If you have been trying to work out what Codex cloud pricing means for a team, that is the thing to understand first, and this post walks through the rest: what each plan includes, what the API-key path costs and gives up, and how the Linear, GitHub, and Slack triggers behave once you are in.
Checked in early September 2026 against OpenAI's own pricing, models, and integration docs, which are linked throughout. OpenAI revises these pages often.
Sinatra is our product, and it sits in an odd spot for this topic. It competes with Codex cloud on the Linear-to-pull-request loop, and it also accepts a ChatGPT subscription as a credential, so we run on this plan structure ourselves. We will say where that colors the reading.
Codex the agent, Codex the CLI, and Codex cloud
The name covers three things. The Codex CLI runs on your laptop. The IDE extension runs in your editor. Codex cloud runs tasks on OpenAI's infrastructure in an isolated environment with your repository loaded, and it is the half that works while you are somewhere else. The pricing rules differ between them, and most of the confusion online comes from applying one half's rules to the other.
This post is about the cloud half.
Codex cloud pricing on each ChatGPT plan
Codex is included across the ChatGPT lineup, per OpenAI's pricing page:
| Plan | Price | Notes |
|---|---|---|
| Free | $0 | Limited usage |
| Go | $8/mo | |
| Plus | $20/mo | The tier most individuals compare against |
| Pro | From $100/mo | Roughly five times Plus's allowance |
| Business | $20 a user a month billed annually, $25 monthly | Two or more users |
| Enterprise and Edu | Contact sales |
Two of the integrations need a paid plan. OpenAI's Linear docs say Codex in Linear is available on paid plans, and the Slack docs list Plus, Pro, Business, Enterprise, and Edu. So the free tier gets you the web app and the CLI, and the tracker triggers start at $20.
The only number OpenAI publishes
The pricing page gives an allowance per plan, and it is a range rather than a count. On Plus, the published estimate for local messages per five hours depends on the model: 10 to 100 on GPT-5.6 Sol, 25 to 200 on GPT-5.6 Terra, and 250 to 2,000 on GPT-5.6 Luna. Pro is documented at five times those figures.
The caveats on that page matter more than the numbers. Cloud chats on ChatGPT plans use GPT-5.6 Sol and, in OpenAI's words, "may use more of your allowance than local messages." And the page says plainly that "these estimates are not fixed message limits; check your usage dashboard for current limits and reset times."
Between them, those lines decide how a team should read the plan. A cloud task draws on the same allowance as your chat and CLI usage, at the heaviest end of it. A ticket that reads a large repository is many messages' worth of tokens, so two that look the same on the board can land very differently on the meter. And when the allowance runs dry, the choice is to wait for the window to reset or move up a tier. There is no overage rate to pay, because there is no rate.
For an individual, this is a decent deal. Plus at $20 is cheap for what a good day of cloud tasks would cost at API prices. For a team it is harder to plan around, because "how many tasks can we run a week" has no stable answer, and the fix for running out is a $100 plan per person rather than a bill that tracks what you used.
The API-key path, and what it turns off
There is a version of Codex with real per-token prices. Sign in with an OpenAI API key instead of ChatGPT and, per the pricing page, you "pay for Codex usage based on API pricing" in the CLI, SDK, or IDE extension, and "model availability follows the API models available to your key." OpenAI recommends it for automation in shared environments like CI.
The same page lists what that path does not include: "No cloud-based features (GitHub code review, Slack, etc.)."
So the two pricing models are not layered. Published per-token pricing and the cloud agent are alternatives. If the reason you wanted Codex was the Linear assignment or the automatic pull-request review, an API key does not get you there, and if the reason you wanted an API key was a bill you could read, the cloud agent does not offer one.
How the triggers behave
Codex cloud is one of the better-connected agents in this market, and the entry points are documented well enough to quote.
Linear is the cleanest. After installing the integration, OpenAI's Linear docs say "you can assign issues to Codex the same way you assign them to teammates," or mention @Codex in a comment thread to delegate work or ask a question. Codex posts updates back to the issue as it works. One detail to know before you promise it to a product manager: when the task finishes, it "posts a summary and a link to the completed chat so you can create a pull request." The pull request is created from the chat, by you, rather than opened automatically on the issue.
GitHub is built around pull requests. The GitHub docs describe mentioning @codex in a pull request to ask questions or start a task, @codex review for a review on demand, and an automatic-reviews setting that posts a review whenever someone opens a PR. The cloud overview page also lists GitHub issues as a starting point, but the integration page itself documents nothing that starts from an issue, so treat issue triggers as something to test on your own repo before relying on them. For issue-driven automation with an audit trail, OpenAI points at the codex-action GitHub Action, which runs in a workflow file you commit and authenticates with an OpenAI API key stored as a repository secret. That is the API-key path, with the API-key rules.
Slack works the way you would hope: mention @Codex in a channel or thread with a prompt, it reacts with 👀, and replies with a link to the chat.
What you cannot change
The model. OpenAI's models page says cloud chats on ChatGPT plans use GPT-5.6 Sol and that "currently, you can't change the default model for Codex cloud chats." The desktop app, CLI, and IDE extension share a config file that can be pointed at other models and providers, but that is the local half again.
If your team has standardized on Claude for hard tickets, Codex cloud cannot help, and running a second agent to get a second model family is the usual outcome.
Estimating a month
There is no per-task number to multiply, so the only method is empirical. Put the agent on real tickets for a week on one Plus seat. Watch the usage dashboard, which is where OpenAI says the current limits and reset times live. Count how often the window runs dry and how often someone waited for it. If that happens on a normal week, the plan you are actually pricing is Pro, and the team cost is $100 a person rather than $20.
The same week tells you something the pricing page cannot, which is how much of your allowance a typical ticket in your repository consumes. A tight monorepo with fast tests eats much less than a sprawling one with a slow install, and the range OpenAI publishes is wide precisely because of that.
Running the same subscription somewhere else
A ChatGPT subscription is a credential, and more than one agent can use it. Sinatra takes a Codex subscription as a connected credential, so tasks you assign from Linear, or start with the Sinatra label on a GitHub issue, run on the OpenAI models that subscription reaches, with no API key involved. Each member can connect their own, so a team where everyone already pays for ChatGPT needs no shared key. Our charge is $20 per workspace member, flat, with nothing added to the model side, and the free tier is 5 tasks a day on your own key or subscription with no card.
Your subscription's usage limits still apply when you run through Sinatra. Sinatra has a web dashboard for setup and task tracking; generally available task starts come from GitHub or Linear, while Slack is a restricted preview. For differences in where work starts, see the comparison page; the alternatives hub puts Codex next to ten other agents.
Common questions
Is Codex cloud free?
The agent is included in the free ChatGPT plan with limited usage, and the web app and CLI work there. The Linear and Slack integrations require a paid plan. There is no per-task charge on any tier; you pay for the plan and get a rate limit.
Can I use my OpenAI API key with Codex cloud?
No. An API key runs the CLI, SDK, and IDE extension at published API prices, and OpenAI's pricing page says that path has no cloud-based features. Cloud tasks, GitHub code review, and Slack all require signing in with a ChatGPT plan.
Does Codex open the pull request when I assign it a Linear issue?
Not directly. Per OpenAI's docs it posts progress to the issue and, when finished, a summary and a link to the completed chat so you can create a pull request from there. The pull request is a step you take.
Can Codex cloud run Claude or another provider's model?
No. Cloud chats run GPT-5.6 Sol and the default cannot be changed. Other providers can be configured for the local CLI and IDE extension through the shared config file.
If you want the assign-a-ticket loop on the ChatGPT subscription you already have, but from a GitHub issue label as well as Linear, you can connect it to Sinatra and try it on the free tier. The pricing page has the whole number.