Codex Gets a Rate Card

OpenAI just turned its coding agent into something that looks a lot more like cloud infrastructure billing than a chat subscription. On April 2, 2026, it changed Codex pricing for ChatGPT Business and new ChatGPT Enterprise plans from per-message estimates to token-based charges. (help.openai.com) That means the meter now runs on three separate things: input tokens, cached input tokens, and output tokens. Input is what you send in, cached input is reused context that is billed more cheaply, and output is what the model writes back. (help.openai.com) The new rate card spells those prices out per 1 million tokens instead of hiding them behind “about one message.” For GPT-5.4, OpenAI lists 62.50 credits for input, 6.250 credits for cached input, and 375 credits for output, which makes generated code far more expensive than repeated context. (help.openai.com) The cheaper models show the same pattern. GPT-5.1-Codex-mini is listed at 6.25 credits for input, 0.625 credits for cached input, and 50 credits for output, while GPT-5.3-Codex and GPT-5.2-Codex both sit at 43.75, 4.375, and 350 credits respectively. (help.openai.com) OpenAI also added one multiplier that can quietly double a bill. The rate card says Fast mode consumes 2 times as many credits, and code review specifically uses GPT-5.3-Codex. (help.openai.com) Codex itself is not just a code autocomplete box. OpenAI describes it as a coding agent that can navigate a repository, edit files, run commands, execute tests, and work in the terminal, integrated development environment, web app, or background cloud sandbox. (help.openai.com) That matters because long-running agents burn tokens differently than short chats do. A tool that reads a repository, carries forward context, runs multiple steps, and writes patches creates a direct incentive to trim prompts, reuse cached context, and route small jobs to smaller models. (help.openai.com; help.openai.com) OpenAI is also reorganizing who pays and how. As of April 2, 2026, ChatGPT Business and ChatGPT Enterprise have two seat types: a standard ChatGPT seat and a Codex-only seat, and the Codex-only seat is usage-based rather than a fixed monthly subscription. (help.openai.com; help.openai.com) In ChatGPT Business, Codex seats require workspace credits for activity, and admins can set automatic recharge, monthly recharge limits, seat-type limits, and even per-user overrides. That is a billing system built for cost control, not a free-form product experiment. (help.openai.com) OpenAI says Codex now averages about $100 to $200 per developer per month, but it also says the number varies widely with model choice, number of running instances, automations, and Fast mode use. Once pricing works like that, software teams start treating prompt length, cache hits, and model selection the way they already treat compute, storage, and database queries. (help.openai.com) The transition is not finished for everyone. OpenAI says existing Plus, Pro, and Enterprise or Education customers remain on a legacy rate card for now and will be migrated to the new token-based structure in the coming weeks, which means the pricing shift has started but is still rolling through the installed base. (help.openai.com)