Codex pricing: what each plan includes and what a team will spend

There is no Codex subscription to buy. OpenAI's Codex pricing page lists Codex and ChatGPT Work in the Free, Go, Plus, Pro, Business, Edu and Enterprise plans [1], so the real question is which ChatGPT plan each person is on. The table puts the pound prices chatgpt.com shows a UK visitor for the individual plans [2] beside the US dollar prices on OpenAI's Codex pricing page [1]. Do not read one column as a conversion of the other: Go is £7 on the UK page and US$8 on the dollar page.

Business seats also come in more than one size, which matters if you budget from the headline price. OpenAI's usage table refers to a Business ($100) seat that gets the Pro 5x estimates, and its help centre separates Business Standard seats from Business Premium seats [1, 3]. The pound prices OpenAI shows UK buyers for ChatGPT Business are set out, with their sources, in our comparison of the ChatGPT and Claude business plans, linked at the foot of this page.

Seat size also decides which models a person can lean on. Pro plans and Business Premium seats can spend their whole allowance on GPT-6 Astra, the model behind GPT-6 Pro. Plus and Business Standard seats get limited Astra use and pay credits beyond it, and Astra can get through an allowance faster than GPT-5.6 Sol [3].

ChatGPT Work is OpenAI's agent for longer, multi-step jobs such as research, analysis, documents and presentations. OpenAI says Work and Codex share usage, so Work inside ChatGPT runs on the same pricing, credits and usage limits as Codex [1, 3]. Voice tasks in the desktop app draw on the same pool [3], and so does ChatGPT for Excel on Plus and Pro [1].

If a finance director reads one part of this page, it should be this. A person's Codex sessions and Work tasks come out of the same allowance, and on credit-based agreements everyone's eligible use draws on one workspace pool, which other eligible ChatGPT features can also use [5]. A Codex budget set for the engineers alone will run short once the rest of the company starts using Work.

A seat does not come with a fixed number of Codex messages. What OpenAI publishes is a range of local messages per five-hour period for each model and plan, and it calls the ranges estimates rather than fixed limits [1]. A sample of that table:

In this sample, the top of each range is eight to ten times the bottom (our arithmetic from OpenAI's table), because a message is not a fixed amount of work. OpenAI says usage depends on the model, the size and complexity of the task, and whether it runs locally or in the cloud. A small script uses a fraction of the allowance; a long task that holds a large codebase in context uses far more. Two tasks that look alike can cost different amounts, so the length of the prompt tells you very little [1].

Local messages and cloud chats also share the allowance, and cloud chats run on GPT-5.6 Sol and may use more of it. Weekly limits may apply as well as the five-hour window. Fast mode and image generation both use the allowance faster, image generation by about 3 to 5 times on average [1].

Hitting the limit does not cut an engineer off mid-task. OpenAI lets the agent finish the turn it is on, subject to fair use. After that, Plus and Pro users can buy credits without changing plan, Business, Edu and Enterprise workspaces on flexible pricing can buy workspace credits, and anyone can run extra local work with an API key at API rates [1].

Credits do two jobs. On credit-based plans they are how eligible usage is paid for, and on other plans they keep work going once the included allowance is used up. OpenAI prices each model in credits per million tokens, with separate rates for input, cached input (text the model has already read and can reuse) and output [1]. Some rows of the table, at Standard speed:

The model column moves the bill more than anything else here. Per million output tokens, GPT-6 Astra is 100 times the rate of GPT-6 Luna and 5 times the rate of GPT-6 Sol (our arithmetic from OpenAI's table), so a default model setting is a budget decision whether or not anyone made it as one. OpenAI gives one per-message reference point: GPT-5.6 usage averages 5 to 30 credits a message [1].

Speed changes the rate too. Fast mode costs 2.5 times the Standard credit rate on GPT-6 Astra, Sol and Luna. There is no separate charge for writing to the cache. GPT-5.6 Sol is on promotional pricing until at least 21 November 2026 [1].

What the pricing page will not give you is a price for a credit. OpenAI says the price of credits and any discounts depend on your plan or agreement, and points Business, Enterprise and Edu buyers to a credit rate card [1, 3]. Take the credit price from your own rate card or order form, not from a third-party article.

On Enterprise and Edu with flexible pricing there are no fixed rate limits: usage scales with credits. Without flexible pricing, Enterprise and Edu seats have the same per-seat limits as Plus for most features [1].

On a credit-based agreement, what people use and what you are billed are two different things. OpenAI's guide for administrators sets out three cases. While committed credits remain, usage draws them down without creating a new charge. When they run out and the account permits overage, usage carries on up to the workspace overage limit at the contracted overage rate. When they run out and it does not, eligible features pause until credits are added [5].

I would spend as long on the controls as on the rates. Administrators can set monthly usage limits per user, per group or across the workspace, a workspace overage limit and usage alerts. OpenAI is explicit that alerts notify without stopping spending, and that an overage setting of "No limit" is not a zero-spend cap [5]. Both are easy to misread when a workspace is first set up. If your Enterprise agreement bills usage in US dollars instead of credits, use the Enterprise USD rate card and your agreement, not the credit arithmetic on this page [1, 5].

OpenAI's one published usage example shows why an average misleads. For first-time ChatGPT Work users in Marketing, annualised credit use came to about US$53 per employee at the median, US$309 on average and US$1,154 at the 90th percentile. The average sits well above the median because a few people use far more [5]. That example is Work in Marketing rather than Codex in engineering, and OpenAI calls it directional, so do not borrow the numbers. Borrow the shape, and plan for your heaviest users rather than the typical one.

Codex will also run on an OpenAI API key instead of a ChatGPT sign-in. Usage is then billed through your OpenAI Platform account at standard API rates, rather than drawn from a plan's allowance [1, 6]. OpenAI suggests it for automation in shared environments such as CI [1].

An API key covers Codex in the CLI, the SDK and the IDE extension, and model choice follows the models your API key can reach. It leaves out the cloud features: Codex cloud, GitHub code review, the Slack integration, ChatGPT Work on the web and the workspace admin features of Business and Enterprise [1]. Data handling changes too. With an API key, usage follows your API organisation's retention and data-sharing settings, not your ChatGPT workspace's roles, retention and residency settings [6]. For a pipeline, that is a fair trade. For an engineer who works in Codex all day, it gives up the cloud features, and on Business or Enterprise the admin controls, that the seat included.

On 14 October 2026 GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on every plan. The OpenAI API is not affected. OpenAI's stated replacements are GPT-6 Sol on Plus, Pro, Business, Enterprise and Edu, and GPT-6 Luna in the desktop app on Free and Go [4].

Two things follow. Anything that names gpt-5.5 has to change before that date: workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks and scripts [4]. And any budget built on GPT-5.5 usage goes out of date the same day. GPT-6 Sol's credit rates are lower than GPT-5.5's on OpenAI's table [1], but its per-message usage differs too, so re-measure after the switch rather than carrying old figures forward.

No published figure, ours included, will tell you what your team will spend, because usage depends on your code, your tasks and your people. The only estimate I would put in front of a budget holder is one built from your own pilot. This method needs two weeks on real repositories and OpenAI's usage dashboard [1].

Context is the lever you control directly, and the first one I would pull. OpenAI's own advice for making limits last is to keep prompts and source material tight, reduce the size of AGENTS.md files and switch off MCP servers you are not using, because each one adds context to every message [1]. We have measured a related effect ourselves, on a different tool. On a Claude Code skill platform at a large international marketing agency, restructuring the skills so each session loads only the context it needs cut the average cost per session by 60%, by our own measurement. The engagement file for that platform records how.

Setting up and rolling out Codex for an engineering team is core 1AYM work, including the hands-on parts of this page: the pilot, the default model, per-user limits and alerts, and the AGENTS.md and MCP setup. We sell it as AI Engineering Transformation, which builds what we call an AI harness around the agent so its changes reach production like every other change. The documented case, the agency platform above, runs on Claude Code rather than Codex.

1AYM is an OpenAI Select Partner, but we build on OpenAI's and Anthropic's tools alike, so the recommendation follows your pilot, not the partnership. Our comparison of Codex and Claude Code for business covers the choice between them, and our AI strategy and roadmap work covers whether to buy seats at all. A fixed-scope statement of work for the pilot and controls can start within a day of signing the scope, and a rollout you have already scoped can be resourced on contract from the associates who work with us, held to the same standard. Book a call below or email tayyeb@1aym.com if that would help.