OpenAI moves Codex from request‑based pricing to per‑token billing

OpenAI changed how Codex gets billed, and the change is more important than it looks. Codex used to feel like a request-metered product — send a task, burn some allowance, move on. (help.openai.com)he model reads, reuses, and writes back. That sounds cleaner — and it is — but it also makes heavy coding sessions much easier to burn through if the model is carrying a lot of context. (help.openai.com) ### What actually changed? On April 2, 2026, OpenAI updated Codex pricing so usage maps to API-style token accounting instead of p(developers.openai.com)sers right away, plus new Enterprise plans. Then on April 23, 2026, OpenAI extended the same system to all existing Enterprise plans, including Edu, Health, Gov, and ChatGPT for Teachers. (help.openai.com) ### What does “token-based” mean here? Codex now charges credits separately for three things — input tokens, cached input tokens, and output tokens. Basically, you pay one rate for what t(help.openai.com)rate for what the model generates. OpenAI says this replaces rough per-message averages with a direct mapping from actual token use to credit consumption. (help.openai.com) ### Which numbers matter most? The spread is wide. GPT-5.5 costs 125 credits per 1M input tokens, 12.5 for cached input, and 750 for output. GPT-5.4 is hal(help.openai.com)edits. So the expensive part is not just “using Codex.” It is using a bigger model, with lots of fresh context, that writes a lot back. (help.openai.com) ### Why does this feel different from request pricing? A request-based system hides a lot of variance. One prompt looks like another even if one is a tiny refactor and the next drags in a giant repo, long c(help.openai.com)penAI even says small scripts may use only a fraction of an allowance, while larger codebases, long-running tasks, and extended sessions that hold more context use significantly more. (developers.openai.com) ### Why are long sessions the trap? Because context is the meter now. If Codex keeps hauling around a large codebase a(help.openai.com)code appears. Cached input is cheaper, which helps, but output is still expensive enough that big rewrites and verbose reasoning can move the bill fast. Think of it less like paying per taxi ride and more like paying for distance, traffic, and waiting time separately. (help.openai.com) ### What about the $100 Pro tier? OpenAI is also running a temporary promo on Pro. Its Codex pricing page says the $100/m(developers.openai.com)he page shows Pro as 10x or 20x more Codex usage than Plus during that window. The catch is that a bigger allowance does not undo token math — it just gives heavy users more room before the meter bites. (developers.openai.com) ### Does this make Codex clearer or scarier? Both. It is clearer because teams can now reason about usage with the same token logic they already use for APIs. OpenAI even says the format is me(help.openai.com)ion, and points users to a Usage panel inside Codex settings. But it is scarier for anyone who liked the fuzziness of message caps, because the new system makes expensive habits impossible to ignore. (help.openai.com) ### So what’s the bottom line? Codex did not suddenly become “more expensive” in one universal way. But OpenAI mo(developers.openai.com)r workflows, and makes model choice matter a lot more in day-to-day coding. If you are a light user, this is mostly a transparency upgrade. If you run long agentic sessions on large repos, this is a budgeting change first and a product change second. (help.openai.com)