OpenAI swaps Codex pricing

OpenAI changed how it charges teams for Codex on April 2, 2026: instead of counting an average “message,” it now bills ChatGPT Business and new ChatGPT Enterprise customers by the number of tokens the coding agent reads, reuses from cache, and writes back. (help.openai.com) A token is a small chunk of text, so this switch turns coding bills into something closer to a cloud-computing meter: longer prompts, bigger codebases, and wordier answers all show up directly on the invoice. (help.openai.com) OpenAI’s new rate card breaks that meter into three parts per 1 million tokens: input tokens, cached input tokens, and output tokens. For GPT-5.4, the listed prices are 62.50 credits for input, 6.250 credits for cached input, and 375 credits for output, which means generated code is far more expensive than reused context. (help.openai.com) The old system hid that detail behind per-message estimates, so a team could send one short bug-fix request and one giant repository-wide refactor request and still think in “messages” instead of raw consumption. OpenAI says the new format replaces those averages with a direct mapping from token usage to credits. (help.openai.com) OpenAI paired the pricing change with a new seat type called a Codex-only seat. Those seats have no fixed monthly fee, give access only to Codex, and bill entirely through workspace credits, which lets a company put a contractor or a small pilot team into the coding tool without buying a full ChatGPT seat. (openai.com) At the same time, OpenAI cut the annual ChatGPT Business seat price from $25 to $20 per seat. The split creates two lanes: a standard seat for broad ChatGPT access with included Codex limits, and a usage-based seat for people who mainly need the coding agent. (openai.com) This also changes who has to care inside a company. Under OpenAI’s flexible pricing rules, Business users can hit included limits and then draw from a shared credit pool, while Enterprise workspaces buy a shared pool and can set spend controls by group through role-based access control. (help.openai.com) Once billing depends on tokens, prompt design stops being just an engineering preference and starts looking like a procurement problem. A verbose system prompt, a large attached codebase, or a habit of asking for long explanations can all raise spend, while cached context is priced much lower and rewards repeat work on the same code. (help.openai.com) OpenAI is openly pitching that tradeoff as easier to understand. Its April 2 product post says Codex-only seats have no rate limits and that token billing gives teams a clearer view of how usage turns into spend across budgets, workflows, and teams. (openai.com) The catch is that “clearer” does not mean “more predictable” for every buyer. OpenAI says Codex costs about $100 to $200 per developer per month on average, but it also says the range varies a lot by model, number of running instances, automations, and use of fast mode, which doubles credit consumption. (help.openai.com) So the practical result is simple: under message pricing, buyers mostly asked how many requests a developer could send; under token pricing, they have to ask how much code the model reads, how often context is reused, and how long the model’s answers are. That is a more precise meter, but it pushes the work of optimization onto the customer. (help.openai.com)