OpenAI Codex usage limits, credits and plans on ChatGPT Business and Enterprise, explained

On April 2, 2026 Codex pricing past the included allowance moved from per-message rates to credits aligned with API token usage⁴¹⁶. On August 24, 2026 ChatGPT Business gained a Premium seat with five times the usage of a Standard seat and no five-hour limit, and a cap of 200 paid seats⁴⁵. A Standard seat is $25 a month, or $20 billed annually, with a minimum of two users²; a Premium seat is $125, or $100 billed annually³. Codex is included in every ChatGPT plan¹, both Business seat types include it³, and local messages and cloud chats share one allowance²⁷.

The allowance is published as estimated local messages per five-hour period by model, with the caveat stated plainly: “these estimates are not fixed message limits"¹. For a Standard Business seat the estimates run from 5 to 45 messages on GPT-6 Astra up to 350 to 3,000 on GPT-6 Luna; a Premium seat uses the five-times figures¹. OpenAI never puts a token figure on a window, only those estimates.

Usage limits and workspace credits

Two units, and only one of them is money:

• A usage limit is the allowance inside a seat: a roughly five-hour window and a weekly window¹². Standard seats carry the five-hour limit; Premium seats do not, and reset weekly⁵⁴. On the personal plans the weekly period runs seven days from a member’s first request¹²; the usage dashboard shows a workspace’s current limits and reset times¹. What a message costs depends on the model and the size of the task: “small scripts or routine functions may consume only a fraction of your allowance"¹, cloud chats “may use more of your allowance than local messages"¹, and image generations use included limits three to five times faster than similar turns¹. Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents draw from one shared allowance and credit pool where the plan offers them¹³. When a member reaches a limit during an active turn, the agent finishes that turn, subject to fair use¹. OpenAI Support does not reset usage limits¹³, and the paid weekly reset offered on Plus and Pro is not available on Business or Enterprise¹².
• A workspace credit is what a member spends to keep working after the included allowance is consumed¹. On Business, only workspace owners buy credits, in Workspace settings under Billing¹⁰; the credits are pooled across all users and seat types⁹; each member’s included usage is consumed first, and only then does eligible activity draw from the pool¹⁰. Credits are valid for 12 months after purchase⁹, are non-refundable and non-transferable, and under the generic terms expire one year after purchase or issuance unless otherwise specified¹⁹. With no credits in the pool, the feature is blocked on Business and paused on Enterprise unless owners enable overages⁹. On Enterprise, credits are a contract-level shared pool bought through the account team, with allocation and expiration set in the Order Form⁹.

Usage inside the windows is not metered in dollars. Past them, it is metered in credits per million tokens⁶.

What a credit buys

Since April 2, 2026 the rate card for Business and Enterprise credit-based pricing is expressed in credits per million tokens, by input, cached input and output⁶:

Credits per million tokens⁶. GPT-5.6 Sol’s rate is promotional and available at least through November 21, 2026¹; GPT-5.5 was retired from Codex with ChatGPT sign-in on October 14, 2026, leaving the OpenAI API unaffected¹. The formula is total credits equals input tokens over one million times the input rate, plus cached input tokens over one million times the cached rate, plus output tokens over one million times the output rate⁶, and those rates apply to local tasks, cloud tasks, automations, code review and delegated workers alike⁶; auto-review safety checks are free on a ChatGPT sign-in and count against nothing⁶. Reasoning effort does not change a model’s rate, only how many tokens a task uses⁶; Fast mode uses 2 times the standard rate where available, Ultrafast on GPT-6 Astra and GPT-6.1 Sol uses 6 times, and speed configurations increase credit consumption on every applicable model¹. A typical Codex task on GPT-6.1 Sol “may consume between 2 and 15 credits per task"⁶. On ChatGPT sign-in no long-context multiplier applies to any Codex model, even above 272K input tokens⁸.

What a credit costs in dollars depends on the plan or agreement¹, and no page cited here prices a Business credit. The Enterprise token-based rate card bills GPT-6 Astra at $10.00 per million input tokens and GPT-5.6 Sol at $4.00⁸. Divide those by the credit rates above and each comes out at $0.04 per credit. That is a ratio of two published rate cards, not a price OpenAI quotes: there is “no single credit or dollar equivalent” for included usage⁵, and OpenAI’s own guidance is to use your agreement’s rates and “not assume a generic credit-to-dollar conversion"¹¹.

Which plan you are on

Enterprise and Edu plans without flexible pricing have the same per-seat usage limits as Plus for most features¹. Individual Free, Go, Plus and Pro plans are not covered on this page.

How the seat model changed in 2026

Under the per-seat windows, one member’s heavy week is one member’s problem, and one member’s quiet week is nobody’s gain. The changes this year repriced what happens past the window and, on Business, added a second seat type, without changing the shape.

• April 2, 2026. Codex pricing past the allowance aligned with API token usage instead of per-message pricing⁴, with new Enterprise workspaces on token-based rates from that date¹⁶. Codex-only seats on flexible, credit-based pricing arrived on Business and Enterprise, and the price of subscription-based seats fell by $5 a month⁴.
• June 18, 2026. Enterprise gained monthly credit limits by workspace, group and user, usage alerts, overage limits, a Spend Controls API, a Cost API and credit analytics by user, product and model¹⁸.
• June 24, 2026. Codex-only seats closed to new Business workspaces¹⁵.
• July 6, 2026. Workspace Agent runs moved to token-based credit pricing⁴, and on July 9 ChatGPT Work launched with the same usage structure as Codex²⁸.
• August 24, 2026. Premium seats arrived on Business, mixable with Standard seats in one workspace, and the plan was capped at 200 paid seats⁴⁵.

Three details that matter more than the dates:

1. On Business, credits are optional, and nothing draws from an empty pool. Until an owner buys credits there is nothing past the window, and the feature is blocked⁹. Automatic reload is off until it is enabled, and a blank monthly recharge limit allows unlimited automatic purchases each month¹⁰.
2. On Enterprise, overage is on by default with no cap. OpenAI does not set a workspace overage limit; setting it to zero is what stops the workspace continuing into overage¹¹.
3. The window belongs to the seat. Included usage is evaluated for the seat assigned to each member¹⁰. A member at the five-hour or weekly wall waits for the reset or draws on pooled credits¹; another member’s unused window is not theirs, and support will not reset it¹³.

Why this quarter may look different

A workspace that fit inside per-message pricing in March is on credits per million tokens now⁴¹⁶, with the default model, speed setting and reasoning effort deciding how fast those credits go¹⁶. OpenAI now publishes its own average for the bill’s size: “On average, Codex costs approximately $100–$200 per developer per month"⁶.

There is no published breakdown of who spends what, only the tools for finding your own: user leaderboards in Codex analytics⁴ and, on Enterprise, the instruction to “identify top users and emerging credit usage patterns"¹⁸. On a per-seat window, the members in those leaderboards are exactly the set who reach the wall.

What actually drains an allowance

Model choice first: 250 credits per million input tokens on GPT-6 Astra against 2.5 on GPT-6 Luna⁶. Then speed: Fast mode multiplies a model’s rate by 2 where available, Ultrafast multiplies Astra’s by 6, and speed settings raise consumption on every applicable model¹. Then task shape: cloud chats may use more of the allowance than local messages¹, reasoning effort raises tokens without changing the rate⁶, image generations burn included limits three to five times faster than similar turns¹, and code review counts when Codex runs it through GitHub¹. Voice in the desktop app uses the Codex budget at $0.05 a minute¹. Fast Mode is enabled by default where admins have not overridden it¹³.

“These estimates are not fixed message limits; check your usage dashboard for current limits and reset times.” OpenAI’s pricing page¹

What an administrator can do

Each control below does one thing and stops at a definite point.

• Workspace analytics (Business). Usage by member, including seat type, credits spent and messages sent, with a Codex view and windows of seven days, one month, six months, twelve months or custom⁴. Owners can view remaining credits and download usage reports under Billing⁹. Where it stops: it shows. It does not decide.
• Admin Console analytics (Enterprise). Codex usage, active users, leaderboards, credits and exportable activity, up to 120 days of history refreshed every one to six hours¹⁷; credit spend broken down by user, product and model, with a Cost API¹⁸; CSV exports by user, GPT and project, with data refreshed every one to 24 hours and typically within six to twelve²⁴. Usage-limit figures are close to real time while other analytics are delayed¹¹. Where it stops: the same place.
• Monthly credit limits (Business). Limits per seat type, Standard and Premium, plus per-user overrides; by default no seat and no user has a limit¹⁰. Owners and Admins set spend controls; Members cannot²⁶. Where it stops: a limit decides when a member stops drawing on the pool. It does not move another member’s unused window.
• Monthly credit or dollar limits (Enterprise). A workspace default, group defaults and user overrides, applied in the order user override, highest applicable group default, then workspace default, on a monthly UTC period or one aligned to the billing cycle¹¹. At a member’s limit, additional eligible usage is prevented, though a task may finish slightly over¹¹. Members can request more credits with context for an admin to review¹⁸. A Spend Controls API automates the limits¹¹. Where it stops: the same place a Business limit stops.
• Overage limit and usage alerts (Enterprise). Overage above the committed pool is allowed by default with no cap; an owner can set a limit, and zero prevents overage entirely¹¹. Alerts on a threshold, on exhausting the committed credits or on the overage limit notify but never block¹¹. Where it stops: an overage limit stops the workspace, everyone at once.
• Enabling Codex. Managed workspaces control local Codex use and cloud tasks separately¹³, with Default, On and Off settings per member or group through custom roles²¹. Admins set the default model, reasoning level and speed, and whether Fast Mode is available¹³. Where it stops: a cheaper default slows the meter for everyone equally. It does not change who runs out first.
• Closing the side doors. Managed configuration can force the login method and restrict users to a specific workspace, so a developer cannot step around the workspace with a personal plan or an API key²⁰. API-key usage is billed through the Platform account at standard API rates, outside the ChatGPT plan²⁰, under the Platform’s own project-level spend limits²⁵. Where it stops: it keeps the spend on one bill. It does not shape it.
• The APIs. The Compliance API covers Codex usage from local clients and cloud¹³ and is “not a productivity dashboard"²²; the Codex Enterprise Analytics API returns aggregated usage and activity metrics for a workspace²³; the Spend Controls API reads and writes usage limits¹¹. Where it stops: they read usage and write ceilings. There is no call that moves allowance.

Five things to check before the next reset

1. Decide whether the pool has anything in it. On Business, nothing is past the window until an owner buys credits, and the feature is blocked when the pool is empty⁹¹⁰. On Enterprise, the opposite default applies: overage runs with no cap until someone sets one¹¹.
2. Set the limits that default to none. Business seat-type and per-user monthly credit limits are unset by default¹⁰; so is the Enterprise overage limit¹¹. A blank automatic-reload monthly limit allows unlimited reloads¹⁰.
3. Check the default model, speed and reasoning. Fast Mode is on by default¹³ and multiplies the rate by 2 where available¹; GPT-5.6 Sol’s promotional rate runs at least through November 21, 2026¹; GPT-5.5 was retired from Codex on ChatGPT sign-in on October 14, 2026¹, so any default still pointed at it has already moved.
4. Look at the distribution, not the average. Workspace analytics lists credits spent by member⁴, Enterprise credit analytics breaks spend down by user, product and model¹⁸, and both export⁹²⁴. The question is which seats sit on the leaderboard and which seats sit on an untouched window.
5. Close the side doors before you tighten the front one. Force the login method and workspace through managed configuration²⁰, and set the Platform project limits for anything that runs on an API key²⁵. A per-member limit on the workspace is only a limit if the workspace is where the usage happens.

The part nobody has solved

Every native control, the credit purchase, the seat-type and per-member credit limits, the overage limit, the monthly budget, the model and speed defaults, is a ceiling. Ceilings stop spending, and none of them moves an unused window to the member who ran out. Business credits are already pooled⁹, but the pool sits behind the per-seat windows: a member at the wall draws on shared credits while another seat’s window resets unused, and the only lever on the heavy member is a cap. Both results are true in the same workspace, in the same week: a blocked engineer on one seat, an allowance expiring on the next. The same is true of the other vendors’ controls, which ship their own ceilings.

We are building Spillway Ops for that missing route: allowance that would otherwise reset unused flows to the members who ran out, so spend falls without anyone losing a tool. It reads usage metadata and sets per-user limits through vendor admin APIs with no repository access, no prompt contents, no code. The Codex connection is planned. On Enterprise, the Spend Controls API’s usage-limit reads and writes¹¹ and the Analytics API’s aggregated usage²³ are the documented paths. If the first invoice on token-based credits is about to be a conversation at your company, we would genuinely like to hear how it goes.