OpenAI posts Codex pricing

OpenAI has posted a detailed Codex rate card that puts hard numbers on how its coding product is billed across consumer, business and government-linked plans. The Help Center page, updated May 19, says Codex now uses token-based pricing for Plus, Pro, Business, Enterprise, Edu, Health and Gov customers, replacing older per-message estimates. The move comes as coding-assistant vendors compete more openly on price and infrastructure. Cursor said on May 19 that its new Composer 2.5 model is aimed at long-running coding tasks at lower cost, while OpenAI has separately said Stargate remains its long-term compute program even as it works through partners and cloud providers. (help.openai.com) The pricing page does not announce a new subscription tier. It does something more basic: it shows how OpenAI wants customers to measure usage, model by model, in credits per million tokens. ### What exactly did OpenAI publish? OpenAI’s Help Center says the current Codex structure applies to new and existing ChatGPT Plus and Pro users, new and existing ChatGPT Business users, and new and existing Enterprise, Edu, Gov, Health and ChatGPT for Teachers customers. (indianexpress.com) The company said a small subset of Enterprise customers remains on a legacy rate card. April 2, 2026, is the date OpenAI says it shifted Plus, Pro, Business and new Enterprise plans from per-message pricing to API-aligned token usage. (help.openai.com) April 23 is the date it says the same update was extended to all existing Enterprise plans, including Edu, Health and Gov. ### Which numbers stand out on the new card? GPT-5.5 is listed at 125 credits per million input tokens, 12.50 credits per million cached input tokens and 750 credits per million output tokens. (help.openai.com) GPT-5.4 is listed at 62.50, 6.250 and 375 credits respectively, while GPT-5.3-Codex is listed at 43.75, 4.375 and 350 credits. OpenAI also says fast mode consumes credits at a higher rate for supported models. (help.openai.com) The page says code review uses GPT-5.3-Codex, and it describes GPT-5.3-Codex-Spark as a research preview whose credit rates are not final. ### Why does the pricing page matter to developers? OpenAI says the token-based format gives users “a clearer view” of how input, cached input and output affect credit consumption. (help.openai.com) The page also says actual usage depends on the token mix in each task, not a flat per-message estimate. The Help Center adds one planning figure for teams: Codex costs about $100 to $200 per developer per month on average, with wide variation depending on model choice, automations, fast mode and the number of instances running. (help.openai.com) Users can monitor remaining credit in the Codex settings usage panel, according to the page. ### What are rivals doing at the same time? Cursor said Composer 2.5 was launched for long-running coding tasks and improved reliability in following complex instructions. (help.openai.com) The Indian Express, citing Cursor, said the model’s changes came from scaled training, more complex reinforcement-learning environments and new learning methods. May 18 was the date Cursor said it introduced Composer 2.5, and the company said it was doubling included usage for a week. The Indian Express also reported that Cursor said it is working with SpaceXAI on a larger model trained from scratch using 10 times more total compute from the Colossus 2 supercomputer. (indianexpress.com) ### Where does Stargate fit into this? OpenAI said on April 29 that Stargate is its “long-term effort” to build compute infrastructure and that it had already surpassed the 10-gigawatt U.S. infrastructure milestone it set for 2029, with more than 3GW added in the prior 90 days. The company said financing models and partnership structures may evolve, but that capacity coming online at scale is what matters. (indianexpress.com) CNBC reported on April 15 that OpenAI pulled back from a direct Norway offtake arrangement with Nscale and was discussing renting capacity from Microsoft instead. OpenAI told CNBC that using Microsoft for Norway compute made more financial sense and fit within existing contracted spending. May 19’s Codex rate card remains live on OpenAI’s Help Center, and OpenAI’s infrastructure page says the company and its partners are evaluating additional U.S. data-center sites beyond the initial 10GW goal. (openai.com) Cursor, for its part, said its next larger model is being trained with SpaceXAI. (help.openai.com) (cnbc.com)