Guide OpenAI · Chapter 7

    Which levers move the AI bill at OpenAI.

    ChatGPT runs on seats per workspace, the API on tokens per project with a hard spend limit, Codex on allowances. Which lever moves which meter is explained here.

    As of: October 2026 · Editor: Marcel Haas

    • ChatGPT Business
    • ChatGPT Enterprise
    • API projects
    • Token
    • Codex
    • Agents API

    Cost levers

    Which meter, which lever.

    For every meter that moves the bill: what drives it and which lever you hold. Mechanics without figures; the amounts are in the proposal.

    1. 01

      Seats per workspace

      What moves the meter: The number of seats and the tier, Business or Enterprise; apps, Company Knowledge and Codex belong to both, dots run through Business Premium and as an Enterprise beta.

      Your lever: Assign role by role after the pilot, through groups and in Enterprise through SCIM; whoever spends the day in email, documents and meetings is often better served by Copilot.

    2. 02

      Tokens per project

      What moves the meter: Input, output, cache and thinking steps per model; the size of the context per call, reasoning models and retries when an output does not fit.

      Your lever: The model per step, a small one for classification and short texts at volume, instructions in the cache, structured output and a hard spend limit per project.

    3. 03

      Codex allowances

      What moves the meter: The tasks per person against the seat's allowance, cloud runs on machines at OpenAI; allowances reset, extra capacity stays extra.

      Your lever: Roles, environments and repositories the administration enables; the consumption per developer in the pilot before further seats get Codex access.

    4. 04

      Runtime on the Agents API

      What moves the meter: Sessions, sandboxes and the hosted browser as long as the agent works, plus the model's tokens; a task in a loop costs until someone stops it.

      Your lever: An end condition per task, context compaction, questions back before actions with consequences, and the Agents SDK in your own application where no sandbox is needed.

    5. 05

      Usage limits of a dot

      What moves the meter: A dot keeps working towards goals on its cloud computer and consumes the plan's usage limits, even when nobody is in the chat.

      Your lever: One dot per person with one workflow, a cost limit and working hours; the administration enables it and reads the usage in the workspace.

    6. 06

      Storage of file search

      What moves the meter: The index of documents the Responses API keeps for file search as long as it exists, and the tokens for every re-index.

      Your lever: Index only the documents the application needs, create indexes with an expiry, and use apps with the person's rights where the workspace already reaches the source.

    7. 07

      Enterprise features

      What moves the meter: SCIM, role-based access, the Compliance API, the Connector Registry and the data controls of the Enterprise tier come with that tier's price per seat.

      Your lever: Enterprise when data protection and IT demand these features, Business for the pilot group before; the decision is in the data boundaries and security chapters.

    8. 08

      Operating time

      What moves the meter: The person who maintains instructions, tools and test set, answers approvals and reads traces, plus updates and the review of new apps and MCP servers.

      Your lever: A named person with you or our Support & Operations team; count the hours in the pilot and record them in the sheet per application.

    As of October 2026, derived from the dated facts of this chapter and the anatomy of the technology page. Codex allowances, the Agents API runtime, what a dot consumes per task and the storage of file search name the mechanism without a dated Stand; the amounts are in the dated articles and in the proposal, never on this page.

    01

    Seats, tokens and allowances run as separate meters.

    At the end of the month the head of finance finds three items that all come from OpenAI. The workspace bills per seat, a fixed amount per person and month, in the Business or Enterprise tier. The API bills per project by tokens, on the platform account, with a spend limit the development team has set. And Codex comes with the plan, with allowances per person that reset at fixed intervals. Added to that are two meters that are easy to overlook: a dot that keeps working towards goals on its cloud computer and consumes the plan's usage limits while doing so, and an agent on the Agents API whose runtime OpenAI operates as long as it works. A seat is fixed, tokens move with usage, allowances reset, runtime ticks per session. Whoever wants to understand the bill first asks which of these meters runs for which work, and only then how high it stands.

    02

    The workspace pays per seat, and the plan tier decides what comes with it.

    Business and Enterprise give a team a workspace with sign-in through SAML SSO; apps, Company Knowledge and Codex belong to both tiers. Enterprise adds user provisioning through SCIM, role-based access, the Compliance API for audit and eDiscovery, the Connector Registry and the features chapter 8 describes. Dots run in Switzerland through Business Premium and as an Enterprise beta after the administration enables them. The lever lies in the assignment and in the tier: a seat applies per person, you grant it through groups, and whoever counts research, drafts, analyses and code as daily work gets the value of the seat out day after day; whoever rarely works that way often manages without a seat. Many organisations run ChatGPT and Microsoft Copilot side by side, for different groups, and the platform choice chapter says which group belongs where. For the pilot the group is enough; after that you decide role by role with the usage from the workspace.

    03

    The API counts tokens per project, with key, expiry date and spend limit.

    Every application receives a project of its own on the platform, and per project you set keys with an expiry date, a hard spend limit and the data region at creation. The model bills by tokens: input with instruction, history and documents, output with the answer, for reasoning models also the thinking steps, and repeated input from the cache more cheaply. The rate depends on the model, and the models chapter says which model gets which step: a small one for classification, extraction and short texts at volume, a large one for analyses with references and coding. File search in the Responses API keeps an index of your documents, and this storage is a meter that stays as long as the index exists. The levers: the model per step, the size of the context per call, instructions in the cache, structured output that saves retries, an index only for the documents the application needs, and the spend limit, deliberately tight for the first project. The platform dashboard shows the usage per project.

    04

    Codex comes with the plan, and allowances reset.

    A developer hands Codex a ticket in the IDE, and the work counts against the allowance of his seat; in the cloud every task runs on a machine at OpenAI, and that counts too. Allowances reset at fixed intervals, and extra capacity that OpenAI grants for a while stays extra and never becomes a budget line. In Enterprise workspaces Codex Cloud starts switched off: the administration enables use through roles, defines environments, secrets and internet access, and distributes rules for local runs in a requirements.toml file. With that it also decides who moves the meter. What a developer consumes per week we read in the pilot, on real tickets, and from that come the number of seats with Codex access and the rule which tasks go to the cloud and which stay local.

    05

    An agent on the Agents API pays for its runtime, not only for tokens.

    On the Agents API OpenAI runs the agent: sessions that persist, sandboxes, context compaction, since September 2026 also a browser hosted by OpenAI. The meter consists of the model's tokens and of the runtime of session, sandbox and browser as long as the agent works; a task stuck in a loop costs until someone stops it. A dot in ChatGPT keeps working towards goals on its cloud computer and consumes the plan's usage limits while doing so, even when nobody is in the chat; what a task consumes there the vendor does not state on the pages we cite, so the pilot measures it. The Agents SDK in your own application pays the tokens and your own compute. The levers: a clear end condition per task, working hours for schedules, context compaction, a cost limit per dot, and questions back before actions with consequences, which stop a loop before it gets expensive.

    06

    Cost per case is the figure that counts, and the pilot delivers it.

    One task, one pilot group, real cases, and at the end a sheet with one figure per case: share of the seats, tokens, runtime and the hours of the person who maintains the agent, divided by the cases it has handled. You set this figure against what the case costs today, and from that comes the decision whether the agent gets more people. Before that come the limits: the spend limit per project, the cost limit per dot, the roles for Codex. The proposal names the frame for build and operation, the dated articles name plan and list prices with their Stand, and we check both again every quarter because OpenAI changes plans, models and counting rules. What the sheet per application records:

    • Meters: which seats, which project with region and spend limit, which allowances and which runtime run for this application, with the Stand at the vendor.
    • Drivers: what moves the meter, from the number of people through context size and model per step to schedules and loops.
    • Limits: the spend limit per project, the cost limit per dot, the roles for Codex, and the person who receives the alert and acts.
    • Cost per case: the value from the pilot, compared with today's workflow, and the date of the next review.

    What you decide.

    Business or Enterprise for the workspace?

    Options

    1. 1Business, with SAML SSO, apps, Company Knowledge and Codex
    2. 2Enterprise, with SCIM, role-based access, the Compliance API and the Connector Registry
    3. 3Start with Business and move when data protection or IT demand the Enterprise features

    Our recommendation: Business as long as a small pilot group works with SAML SSO, apps, Company Knowledge and Codex and the data protection officer is content with the plan's data controls. Enterprise as soon as accounts should disappear with the leaver through SCIM, the Compliance API is needed for audit and eDiscovery, the Connector Registry should list the allowed apps, or dots run as an Enterprise beta; the data boundaries and security chapters name the reasons.

    What we advise against: Enterprise for the pilot because it is the bigger tier, and equally putting the whole company on Business without SCIM: accounts of people who have left then remain.

    Seats for everyone or for some, with Copilot alongside?

    Options

    1. 1A seat for every person
    2. 2Seats for the roles that research, write, analyse or code daily, Copilot for the others
    3. 3A pilot group, then role by role by usage

    Our recommendation: The pilot group first, then role by role. Whoever researches, writes drafts, analyses data or builds code daily gets the seat; whoever spends the day in emails, documents and meetings is often better served by Microsoft Copilot, and many organisations run both for different groups. You grant seats through groups, in Enterprise through SCIM, and the usage in the workspace delivers the figures for every further group.

    What we advise against: Seats by headcount before the pilot, and equally prescribing the same tool for everyone. The role decides which tool carries the day.

    One project per application or one project for everything?

    Options

    1. 1One project for the whole organisation, with one key
    2. 2One project per application, with its own key, spend limit and data region
    3. 3One project per application and stage: test and production apart

    Our recommendation: One project per application and stage. Each with keys that expire and are rotated from the vault, a hard spend limit, the data region at creation and, where needed, Zero Data Retention after approval. The dashboard shows the usage per project, the alert goes to the person in the sheet, and a compromised key hits one application, not all.

    What we advise against: One key for everything. Nobody sees which application causes the usage, the limit brakes everyone at once, and a leaked key opens everything.

    Agents API, Agents SDK or Responses API, with an eye on the bill?

    Options

    1. 1The Agents API, on which OpenAI runs the agent with sessions, sandbox and browser
    2. 2The Agents SDK in your application, with tokens and your own compute
    3. 3The Responses API for direct calls without an agent runtime

    Our recommendation: By the task, and only then by the meter. Long-running tasks with files, sandbox and browser belong on the Agents API, and for that you pay the runtime; a workflow with its own tools inside your application belongs in the Agents SDK, with tokens and your compute; a single step with a model call in the Responses API. The pilot reads per case what each variant consumes, and the data region from chapter 6 is the second limit, because the Agents API is not listed in the Europe region.

    What we advise against: Choosing the runtime because it is convenient, without reading what a task consumes there and without an end condition for the task.

    Status and sources.

    As of: October 2026

    • As of 12 October 2026

      ChatGPT Enterprise brings SCIM, role-based access and the Compliance API.

      Business and Enterprise give teams a workspace with sign-in through SAML SSO; Codex is included in both plans. Enterprise adds user provisioning through SCIM, role-based access, the Compliance API for audit and eDiscovery, and the Connector Registry.

      Source: OpenAI: Business plans
    • As of October 2026

      The data region of the project decides where a model computes.

      You set the data region per project when creating it, together with keys that expire and hard spending limits. The Europe region, which OpenAI labels EEA and Switzerland, stores and processes in the region for the listed endpoints and requires a retention option approved by OpenAI; the Agents API is not listed there.

      Source: OpenAI: Data controls in the OpenAI platform
    • Since 22 July 2026

      Hard spend limits per project, keys with an expiry date.

      According to the platform changelog, hard spend limits per project have existed since 22 July 2026 and project keys have carried an expiry date since 10 September 2026. Both are set per project, together with the data region at creation.

      Source: OpenAI: API changelog
    • As of October 2026

      Codex Cloud starts switched off in Enterprise workspaces.

      Administrators enable use through roles, define environments, secrets and internet access, and distribute rules for local runs in a requirements.toml file. Codex Cloud respects the repository permissions of the source system.

      Source: OpenAI: Codex admin rollout guide
    • Since 29 September 2026

      OpenAI brings dots.

      Dots keep working towards goals in ChatGPT on their own cloud computer and are reachable by chat, voice, Slack or Teams. In Switzerland through Business Premium and as an Enterprise beta, not for Pro accounts at launch.

      Source: OpenAI Help Center: Getting started with your dot
    • Since 10 September 2026

      The Agents API is in public beta, since 29 September 2026 with computer use.

      OpenAI has run the Agents API as a public beta since 10 September 2026: sessions that persist, sandboxes, context compaction and collaboration between several agents on the runtime OpenAI itself runs Codex on. Since 29 September 2026 an agent also operates a browser hosted by OpenAI through it.

      Source: OpenAI: API changelog
    • As of October 2026

      Agents API, Agents SDK or Responses API.

      OpenAI matches the Agents API to long-running tasks where OpenAI runs the agent, the Agents SDK to agents with their own tools and workflows inside your application, and the Responses API to direct model calls or an agent built from scratch. ChatKit provides an embedded chat interface.

      Source: OpenAI: Agents

    Frequently asked questions.

    Is Codex included in the plan, or do we pay extra?

    Codex belongs to ChatGPT Business and Enterprise and counts against the allowances of the seat, which reset at fixed intervals; in Enterprise workspaces Codex Cloud starts switched off, and the administration enables it through roles. What a person consumes per week and whether tasks go to the cloud we read in the pilot on real tickets. Extra capacity that OpenAI grants for a while we do not plan on.

    What does a dot consume when nobody is in the chat?

    A dot keeps working towards its goals on its cloud computer and consumes the usage limits of the person's plan while doing so, at night too. What a task consumes there OpenAI does not state on the pages we cite; that is why every dot in the pilot gets a workflow, a cost limit and working hours, and the administration reads the usage in the workspace before further people get a dot.

    What happens when a project reaches its spend limit?

    The limit is hard and applies per project, so it hits one application and not all. We set it deliberately tight for the first project, read the usage in the platform dashboard and set it again after the pilot with the measured tokens per case; the person in the application's sheet knows what to do when the limit comes closer. Test and production have separate projects, so that a test run does not brake production.

    Do we need ChatGPT Enterprise to keep costs under control?

    Not for the cost: you grant seats through groups in both tiers, and the API bills per project regardless of the workspace. You need Enterprise for SCIM, role-based access, the Compliance API, the Connector Registry and the data controls that data protection and IT demand; that is a question for the data boundaries and security chapters, and the price per seat follows the tier.

    What does an agent on the Agents API cost compared with the Agents SDK?

    On the Agents API you pay the model's tokens and the runtime OpenAI operates: sessions, sandboxes and the hosted browser, as long as the agent works. In the Agents SDK you pay the tokens to OpenAI and the compute in your own environment. Which variant is cheaper per case depends on how long a task runs and whether it needs sandbox and browser; the pilot measures both, and the data region is the second limit.

    Do we pay twice if a team uses ChatGPT and the API?

    The meters are separate: the workspace bills per seat, the API per project by tokens, and Codex counts against the allowances of the seat. A person with a seat who also builds an application through the API moves both meters, but for different work. The application's sheet records which meter runs for which task, and so the bill looks as planned.

    Where do we find OpenAI's prices and plans?

    On the business plans page and the API price list at OpenAI, in the dated articles below, which name the prices with their Stand, and in the proposal, which records the frame for your application. This page explains the mechanics and therefore stays without amounts; we check plans, models and counting rules every quarter, and the date of the last check is in the article.

    Next chapter · Chapter 8

    How agents and models at OpenAI stay under your control.

    How do agents and models at OpenAI stay under our control?