OpenAI enterprise limits, credits

OpenAI’s enterprise packaging just got a lot more legible. That sounds boring, but it changes how companies buy and govern AI. The old story was basically “buy seats, get broad access, sort out the details later.” The new story is much more granular — which models are included, which ones are effectively unlimited, which features burn shared credits, and how coding work gets metered. ### What actually changed? Three things moved at once. First, OpenAI updated the ChatGPT Enterprise and Edu model-and-limits documentation to spell out current model access, context windows, and the fact that flexible-pricing workspaces should look to a rate card for credit consumption. Second, it expanded credits as a pay-as-you-go add-on beyond plan limits. Third, it published a more explicit Codex rate card tied to token usage instead of rough per-message estimates. (help.openai.com) ### Why does that matter for enterprise buyers? Because “unlimited” was never the whole truth. Enterprise still gets virtually unlimited GPT-5.5 Instant usage, but advanced features now sit inside a clearer economic frame. Thinking models, deep research, image generation, advanced voice, and Codex can draw from credits under flexible pricing. That makes budget planning less about vague adoption and more about what employees are actually doing with the product. (help.openai.com) ### How do Enterprise and Edu work now? Enterprise and Edu workspaces use a shared credit pool bought at the contract level. Users and seat types pull from that pool for advanced features, and admins can set spend controls by group with role-based controls instead of relying on fixed per-seat caps. Once the pool runs out, advanced features pause unless overages are enabled or more credits are purchased through the account team. (help.openai.com) ### How is that different from Business? Business is the hybrid version. Users get per-seat limits first, then can keep going if the workspace has purchased credits. Business and Enterprise also got two seat types as of April 2, 2026 — a standard ChatGPT seat and a Codex-only seat. Edu does not get those seat types, which is an important distinction if you were assuming all org plans were converging on the same packaging. (help.openai.com) ### What’s the big Codex shift? Codex moved to token-based pricing in April 2026. OpenAI says the switch started April 2 for new and existing Plus, Pro, and Business users plus new Enterprise plans, then expanded April 23 to all existing Enterprise plans, including Edu, Health, Gov, and ChatGPT for Teachers. That is a real change in how usage gets understood — less like “messages,” more like API economics inside ChatGPT. ### What do the numbers look like? (help.openai.com) For Codex, GPT-5.5 is listed at 125 credits per million input tokens, 12.5 for cached input, and 750 for output. GPT-5.4 is half that on input and output, and smaller models are cheaper still. OpenAI also says fast mode costs more where supported, and gives a rough average of about $100 to $200 per developer per month, with wide variance based on model choice, automations, and instance count. (help.openai.com) ### Where do flexible credits show up outside enterprise? OpenAI also added credit purchases for ChatGPT Free, Go, Plus, and Pro plans when users hit included limits. Right now those credits work for Codex — for Plus and Pro only — and for ChatGPT for Excel, with Excel flexible pricing taking effect May 5, 2026. Credits are shared across supported features, and eligible users can turn on auto top-up. (help.openai.com) ### So what’s the real takeaway? OpenAI is turning ChatGPT into a bundle of workloads, not a single subscription. That means enterprise buyers will start asking sharper questions — how much coding, how much reasoning, how much agentic spreadsheet work, how much voice. And it means anyone trying to forecast demand for AI infrastructure should care less about logo count and more about whether these workload definitions are becoming concrete enough to budget against. (help.openai.com 1) (help.openai.com 2)