OpenAI publishes per-token credit rates for Codex

OpenAI just made Codex pricing a lot more legible. The big change is simple — instead of estimating usage in fuzzy per-message terms, Codex now shows a published rate card tied directly to tokens. That sounds small, but it fixes a real planning problem. If you run coding agents across a team, you need to know what a long prompt, a cached context window, or a big output will actually cost. OpenAI pushed that change on April 2 for most plans, then finished the rollout to existing Enterprise customers on April 23. (help.openai.com) ### What actually changed? Codex usage is now priced in credits per million input, cached input, and output tokens. OpenAI says this replaces the older per-message framing with a direct mapping to model activity. In plain English, the meter now tracks the thing developers already understand from API billing — tokens in, tokens out, and cheaper cached context. (help.openai.com) ### Who does this apply to? The new rate card covers new and existing ChatGPT Plus and Pro customers, new and existing ChatGPT Business customers, and Enterprise-family plans including Edu, Gov, Health, and ChatGPT for Teachers. OpenAI also says a small subset of Enterprise customers stays on a legacy rate card for now. (help.openai.com) headline rate is GPT-5.5: 125 credits per 1M input tokens, 12.50 credits per 1M cached input tokens, and 750 credits per 1M output tokens. GPT-5.4 is half that on the same pattern — 62.50, 6.250, and 375. GPT-5.4 Mini drops further to 18.75, 1.875, and 113. GPT-5.3-Codex and GPT-5.2 both sit at 43.75 credits for input and(help.openai.com)d research preview, so those rates are not final. (help.openai.com) ### Why is that a bigger deal than it looks? Because this turns Codex from a “how much usage do we think we get?” product into a meterable engineering tool. Teams can now estimate cost the same way they estimate cloud or API spend. If a workflow mostly reuses cached context, the bill looks very different from one that keeps generating long outputs. That make(help.openai.com)ompute you can actually model. (help.openai.com) ### Is this separate from ChatGPT seats? Partly. OpenAI’s April 2 product update introduced Codex-only seats for ChatGPT Business and Enterprise with no fixed seat fee and pay-as-you-go token billing. Teams that want broader ChatGPT access can still buy standard ChatGPT Business seats, and OpenAI cut the annual Business price from $25 to $20 per seat at the (help.openai.com)se, or usage-based Codex access for teams that mainly want the coding agent. (openai.com) ### What does OpenAI say typical spend looks like? OpenAI’s help page says average Codex cost runs about $100 to $200 per developer per month, but with large variance. The drivers are model choice, how many instances people run, whether automations are active, and whether fast mode is on. That caveat matters — a light IDE assistant user and a team running lots of background coding jobs are not buying the same thing. (help.openai.com) ### How does this line up with API pricing? Pretty closely in structure. OpenAI’s API pricing page already breaks models into input, cached input, and output token prices — for GPT-5.5, that is $5, $0.50, and $30 per 1M tokens on the API side. Codex credits are a different billing unit, not a dollar mirror, but the logic is now the same. That consistency is (help.openai.com) and harder to misunderstand. (openai.com) ### Bottom line This is less about a price cut than a pricing translation. OpenAI is telling teams exactly how Codex usage turns into spend — and once that math is visible, Codex starts looking less like a premium seat add-on and more like programmable infrastructure. (help.openai.com)