How Much Does Codex Cost a Team? Seat vs API Pricing, Worked Out for 10 and 50 Engineers (2026)

TL;DR

OpenAI does not publish what Codex costs per developer, so you have to build the number yourself. The seat prices are public on OpenAI's Codex pricing page: ChatGPT Business is $20 per user per month billed annually ($25 monthly), Pro is $100, $200 or $500 a month, and Enterprise is quoted by sales. Usage beyond a seat's allowance runs on credits, metered per million tokens at rates that track API list prices. Our own measurement puts a fully used Codex Pro 5x seat at about $1,200–1,500 a month of API-equivalent work. With the assumptions written out below, a 10-engineer team consumes roughly $4,000–6,200 a month at API rates and a 50-engineer team roughly $20,000–31,000. The seat invoice will show a fraction of that.

This post covers how each billing model meters, what the tokens cost, a worked budget for 10 and 50 engineers, and where the spend goes.

The four ways a team pays for Codex

Codex comes bundled with ChatGPT plans or billed through an API key, and OpenAI's pricing page lists both. The setups meter differently:

Most companies end up with more than one of these. An engineer who outgrows a Business seat buys Pro on a card, the CI pipeline runs Codex on an API key, and the platform team negotiates Enterprise. Each line arrives on a different invoice in a different unit (seats, credits, dollars per token), and none of them says what the work cost.

How credits and limits meter

A seat allowance is a budget of tokens that you never see in dollars. OpenAI expresses it two ways.

On Plus and standard Business seats, the pricing page gives local-message ranges per five-hour period by model: 5–45 messages on GPT-6 Astra, 15–160 on GPT-6.1 Sol, 35–285 on GPT-6.1 Terra and 350–3,000 on GPT-6 Luna. The ranges are wide because a message is not a fixed size. A short question against a small file costs little; a request that drags 150,000 tokens of repository context costs a great deal, and both count as one message. Cloud tasks, the page notes, "may use more of your allowance than local messages."

On workspace plans the window can look different. In the Codex usage we have captured from team and enterprise workspaces, the limit reported as a single window that resets about monthly rather than as a five-hour window plus a weekly one. A monthly window behaves like a prepaid balance: nothing slows an engineer down on day three, so heavy users can drain most of it early and then hit a wall with weeks left before the reset. Five-hour windows spread the pain; a monthly window concentrates it.

Past the allowance, credit-based plans pay in credits. The rate card lists credits per million tokens:

OpenAI does not publish a dollar price per credit. The page says "credit purchase prices and applicable discounts depend on your plan or agreement." Compare the rows with the API price list below and every one is exactly 25 times the dollar rate, so your contract's price per credit sets how credit-billed usage compares with an API key. Speed changes the meter too: fast mode uses 2x credits on purchased credits and 2.5x on included subscription usage, and ultrafast uses 6x and 8x, per the same page.

What the tokens cost at API rates

Under every setup, the unit underneath is the token. Current list prices from OpenAI's API pricing, standard tier, per million tokens, for prompts up to 272K input tokens:

Model choice is the largest lever on that table. Astra costs five times what GPT-6.1 Sol does on input and output, and ten times as much on cached input. Our measurement showed the effect on a seat: after our account's default moved from Astra to GPT-6.1 Sol, one point of the Pro 5x weekly limit went from $3.40 to about $2.70 of API-priced work. Each point covered more tokens, because Sol is cheaper per token, but fewer API dollars. The allowance is not metered strictly in proportion to list price, which is one more reason to measure it rather than infer it.

A worked budget for 10 and 50 engineers

OpenAI publishes no per-developer average for Codex, so this budget starts from our one measured account and adds assumptions you can replace with your own:

• Heavy users (20% of the team) consume what a fully used Codex Pro 5x seat carried in our measurement: $1,200–1,500 a month at API list prices.
• Everyone else consumes $200–400 a month at API rates. This is an assumption, not a measurement; it sits in the same range Anthropic publishes for Claude Code ($150–250).
• List prices, no negotiated discount, annual billing where it applies.

Two cautions go with this table. First, the heavy-user figure comes from a Pro 5x seat used to 100% of its weekly limit; it is a ceiling on one plan, not an average, and a Business seat's allowance is smaller. Heavy users on Business will run out and either wait or spend credits. Second, the gap between the seat rows and the usage rows is not money you save. It is usage the vendor absorbs inside the allowance until someone runs past it, and then it reappears as credits, an upgraded Pro seat, or an API key somebody set up to keep working.

Why the seat price hides your real spend

A $20 seat is the floor of what a Codex user costs, never the total. The usage underneath has a real API-rate value, and that value is the only figure you can compare across tools. If half the team uses Codex on Business seats and the other half uses Claude Code or Cursor, a $20 seat on one side and a per-token bill on the other say nothing about which tool did the same work for less. Pricing every tool's tokens at its API rate puts them on one scale, which is the core move of TokenOps.

A flat seat also hides waste. If a third of a team's tokens are context re-sent on every turn, the seat invoice does not move. The allowance drains faster, engineers hit the limit sooner, and the fix people reach for is a bigger plan.

Where Codex spend goes

Output is the small part. A coding agent re-sends the conversation and the files it has read with each request, so the cost of a turn is mostly the context it carries. Take one request holding 150,000 tokens of context on GPT-6.1 Sol at API list price:

• Context served from cache: 150,000 × $0.10 per million = $0.015
• The same context uncached: 150,000 × $2.00 per million = $0.30, twenty times more

On GPT-6 Astra the same two numbers are $0.15 and $1.50. Across a team, the recurring sources are consistent:

• Context carry: Sessions that run across unrelated tasks keep paying for their history. See context efficiency.
• Re-reads: The agent reads the same files again in later turns, or scans broadly because the prompt was vague.
• Oversized models: Astra as a default for routine edits costs five times what Sol does on the same tokens.
• Fast mode left on: At 2–2.5x credits for fast and 6–8x for ultrafast, a speed setting chosen once multiplies every task after it.

Together these are recoverable token waste: spend you can remove without removing any of the work.

How to get an honest number for your team

1. Put every billing path on one ledger. Business and Pro seats, workspace credits and API keys, priced at API rates so they add up.
2. Attribute it per repo and per workflow, so you can see which work produced the spend. Our guide to attributing AI coding spend covers how.
3. Separate waste from leverage. The aim is the same work at a lower cost, not less work.
4. Watch the trend. A default-model change moved what our seat delivered by about a fifth, with no change to the invoice. Track it with the TokenOps metrics that show it.

Promptster Teams gives you the real number

Promptster Teams attributes AI coding spend across Claude Code, Codex, and Cursor by repo and workflow, priced at API rates so seat plans and usage bills land on one ledger. Each engineer sees their own figures; managers see team and squad totals, never a per-person ranking. It separates recoverable waste (carried context, re-reads, oversized models) from spend that shipped work, and tracks the trend. Code and diffs are redacted on the engineer's machine before upload, and it is not developer monitoring.

If the honest answer to "how much does Codex cost us?" is "we don't know," that is the problem it solves.

See how Promptster Teams attributes AI coding spend →

Related: How much does Claude Code cost a team? · How much does Cursor cost a team? · Claude Code vs Codex vs Cursor cost · Claude Team vs Enterprise vs API · What a $100 AI coding plan is worth · What is TokenOps? · Recoverable token waste