OpenAI moves from scale to rationing
OpenAI is quietly shifting from wide-open access to a more controlled, enterprise-oriented model where powerful reasoning features are rationed and smaller fallbacks are formalised. The company’s Codex rate card and release notes show enterprise migrations will be managed directly and that GPT-5.4 mini will act as a fallback when rate limits hit, while usage limits were tightened after Codex reached roughly 3 million weekly users. For buyers, that means performance guarantees will depend on routing rules and fallback behaviour as much as headline model quality. (help.openai.com) (help.openai.com) (technobezz.com)
OpenAI used to sell ChatGPT like an all-you-can-eat buffet, but its newest help pages read more like a utility bill with surge pricing, fallback generators, and account managers deciding who gets moved to the new meter first. On April 2, 2026, OpenAI updated the Codex rate card so new and existing ChatGPT Business customers and new ChatGPT Enterprise customers are billed on token usage, while older Plus, Pro, and Enterprise or Education customers stay on a legacy system until migration in the “upcoming weeks.” (help.openai.com) Codex is OpenAI’s coding agent, which means it does more than answer a question once and stop. The help center says Codex can work in the terminal, inside an integrated development environment, or in the Codex app, and it is included across Free, Go, Plus, Pro, Business, and Enterprise or Education plans with different limits. (help.openai.com) The pricing change matters because token billing turns usage into a moving target. OpenAI says actual Codex credit use now depends on the mix of input tokens, cached input tokens, output tokens, task size, model choice, and reasoning requirements, which is a much less predictable promise than the old “average credits per message or pull request” card. (help.openai.com) At the same time, OpenAI is making fallback behavior an official product feature instead of an invisible emergency patch. In the April 2026 release notes, the company says GPT-5.4 mini will automatically step in for GPT-5.4 Thinking when Plus, Pro, and other paid users hit rate limits. (help.openai.com) That fallback model is not even presented like a normal choice. OpenAI says GPT-5.4 mini will not appear as a selectable model in the model picker, while Enterprise customers can choose to make Auto routing default to GPT-5.4 mini if they prefer. (help.openai.com) So the product now works more like an airline cabin than a single engine. You may buy a ticket labeled “thinking,” but the actual trip can be routed onto a smaller plane called GPT-5.4 mini when traffic is heavy, and the release notes say that switch is designed to preserve access during high usage. (help.openai.com) OpenAI is also separating self-serve customers from managed customers more clearly. The Codex rate card says new and existing Business customers and new Enterprise customers move onto the token-based card now, while customers on existing Enterprise plans are told they will be migrated in the coming weeks, which implies direct account-level handling instead of one universal changeover. (help.openai.com) The timing lines up with a demand spike that OpenAI has now acknowledged in public. Technobezz reported on April 9, 2026 that OpenAI reset Codex usage limits after Codex reached roughly 3 million weekly users, and an OpenAI-linked summary carried by MSN said the milestone was confirmed on April 7, 2026 after rising from 2 million in less than a month. (technobezz.com) (msn.com) That is the quiet shift underneath the marketing. OpenAI is still advertising stronger models, but the practical product is now a bundle of limits, routing rules, seat types, and fallback models, and the Enterprise and Education release notes say Codex beyond baseline access can also use optional flexible pricing tied to workspace credits. (help.openai.com) For companies buying ChatGPT, the question is no longer just “which model are we getting.” The harder question is what happens on the day your team hits the cap: whether work pauses, whether it spills into token-priced overage, or whether it silently drops onto GPT-5.4 mini and keeps going with a different cost and performance profile. (help.openai.com 1) (help.openai.com 2)
