Codex Pricing: What It Actually Costs

Part of the reduce AI coding agent token usage pillar. This page is about what Codex costs; the pillar is about spending less whichever agent you run.

How much does OpenAI Codex cost?

Codex has no price of its own. It is included across ChatGPT plans — Free and Go included — and your plan decides how much of it you get, which is why "how much does Codex cost" has no single number to answer it. What changed in 2026 is the shape of that answer: since April, usage is metered in token credits rather than counted in messages. That distinction matters more than the plan price. Under a message quota, a one-line question and a forty-file refactor cost you the same. Under token credits, they do not, and the gap between them is the whole game. Two developers on the same plan can now burn through it at completely different rates without either of them doing anything unusual. This page covers what you are actually buying, what changed and when, and what moves the number. For the individual plan prices, check OpenAI's pricing page directly rather than any guide including this one — they are rendered live and they move.

Which ChatGPT plans include Codex?

All of them. OpenAI's help centre states that Codex is included across ChatGPT plans, including Free and Go, with usage limits varying by plan. Paid tiers describe it as "expanded Codex usage" (Plus) and "maximum Codex tasks" (Pro). You reach it through four official clients, all signed in with the same ChatGPT account:

the ChatGPT desktop app, in Codex mode
the Codex CLI, which is where most agentic work happens
the Codex IDE extension
Codex on the web

The client you pick does not change your entitlement. It changes how much context the agent pulls in before it starts, which is the part that costs you.

What changed in April 2026?

OpenAI replaced per-message pricing with token-based credit billing across Plus, Pro and Business. The announcement, dated 2 April 2026, put it plainly: "Codex-only seats have no rate limits, and usage is billed on token consumption. This gives you a clearer view of how usage turns into spend." This is the single most important fact about Codex pricing, and it is also why most of the guides ranking for this query are misleading. Anything written before April describes a message-quota product that no longer exists. Anything that still frames the question as "how many messages do I get" is answering about a different product. The same announcement lowered the annual price of ChatGPT Business from $25 to $20 per seat.

What changed again in June?

Pay-as-you-go Codex seats stopped being available to new Business customers on 24 June 2026. OpenAI added the note directly to the April announcement: "Starting today, new Codex pay-as-you-go seats will no longer be available for Business plans. Existing pay-as-you-go seats are not affected." If you are on a grandfathered pay-as-you-go seat, nothing changed for you. If you are evaluating Codex for a team today, the unlimited-rate, pay-only-for-tokens option that most 2026 guides describe is not on the table — you are looking at standard Business seats, which do carry Codex usage limits. This is worth checking before you plan a rollout around a pricing model you read about in a guide dated June or July. Several of the top results for "codex pricing" still describe pay-as-you-go as generally available.

How do token credits actually work?

A credit is consumed by volume of text moved, not by number of requests. Every turn sends the model your prompt plus whatever the agent decided to attach — file contents, command output, tool definitions, prior conversation — and the whole payload is metered. The reply is metered too, at a higher per-token rate, but there is far less of it. The practical consequence is that cost tracks context size, and context size is something the agent decides on your behalf. You type twelve words; the agent reads four files, runs a test, and pulls in three tool manifests. You are billed for all of it. This is why two people on an identical plan can exhaust it at wildly different rates: the difference is not how much they type, it is how much their setup reads before answering. It also explains why the old message quota felt fairer to light users and worse for everyone else. A quota of N messages priced a one-line question exactly like a repository-wide refactor. Token credits price them honestly — which is better, but it removes the illusion of a predictable monthly ceiling unless you actively manage what enters context. Two habits follow from this, and they are worth forming early. First, treat every "read the whole file" as a purchase. Second, treat every connected MCP server as a standing subscription: its manifest ships on every turn whether or not you call a single one of its tools.

Why does this matter more each month?

Because Codex usage is growing faster than most teams' understanding of how it bills. OpenAI reported on 2 April 2026 that more than 2 million builders use Codex every week, that over 9 million paying business users rely on ChatGPT for work, and that the number of Codex users inside ChatGPT Business and Enterprise had grown 6x since January. Adoption at that pace, on a billing model that changed in April and tightened in June, is exactly the setting in which teams discover their cost model was wrong three months after committing to it. The plan you chose in February was priced in messages. It is not any more.

What actually drives a Codex bill?

Input, not output. For nearly every agentic workload the tokens flowing into the model dominate the tokens coming out, usually by around an order of magnitude. That single fact decides where your effort is worth spending. The things that inflate input are boring and consistent across every agent, not just Codex:

Whole-file reads. The agent opens a 2 000-line file to change four lines, and the other 1 996 are billed.
Command output. A test run, an install log or a git diff gets piped back in full, most of it noise.
Tool manifests. Every MCP server you connect ships its tool definitions into context before your first message, whether you use those tools or not.
Re-reads. The same file entering context four times in one session is four times the cost, and it is the failure mode you are least likely to notice.

None of this shows up as a line item. It shows up as your credits draining faster than the work seems to justify.

Is Codex cheaper than Claude Code, Cursor or Copilot?

Not answerable as a plan-price comparison, and that is the honest answer. All four now meter usage in some form, so the sticker price sets a ceiling rather than a cost. What you pay inside that ceiling depends on how much context your agent pulls per task — which is a property of how you work, not of the logo on the tool. If you want the same breakdown for the others, they are here: Claude Code pricing, Cursor pricing and Copilot pricing. The shape of the answer is the same in all four.

How to apply this today

Check your actual consumption before changing plan. Upgrading to buy headroom you are wasting is the most common and most expensive mistake in this category.
Look at the input/output split first. If input dominates by 10x — and it almost certainly does — then output-side settings are not where your bill lives.
Audit what your agent reads before it reasons. Whole-file reads, unfiltered command output and always-on MCP manifests are the three usual suspects, in that order.
Re-check the plan page rather than a guide. Including this one. The token-credit model means the plan price is the least interesting number on it.

If you would rather not do steps 2 and 3 by hand, that is the gap Tokenade fills: it sits between the agent and its tools, replaces whole-file reads with structure-first reads and semantic search, compresses command output, and defers MCP manifests until they are used. It supports Codex natively alongside the other major agents, and its free tier covers a lot of solo work outright.

What goes wrong (anti-patterns)

Reading a pre-April guide. The message-quota framing is obsolete and it will lead you to plan capacity in the wrong unit entirely. Planning a team rollout around pay-as-you-go. Closed to new Business customers since 24 June 2026. Confirm your seat type before you build a budget on it. Optimising output length. Capping response length feels like cost control and moves almost nothing, because agent traffic is overwhelmingly inbound. It is the lever people reach for first and the one that pays least. Treating the plan price as the cost. Under token credits the plan sets a ceiling. Two people on the same tier can have wildly different effective costs, and the difference is entirely in what their agent reads.

See also: