OpenAI changes Codex pricing
OpenAI introduced a new Codex rate card and signalled that legacy pricing will be phased for some customers, underscoring how AI vendors are actively reshaping packaging and limits. That pricing volatility matters to enterprise buyers who need predictable costs—procurement increasingly asks whether AI features will create usage-driven budget surprises. (help.openai.com)(platform.claude.com)
OpenAI just changed how it charges for Codex, its coding agent: on April 2, 2026, it switched many customers from a rough “credits per message” system to “credits per million tokens,” which means billing now tracks how much text the model actually reads and writes. (help.openai.com) That sounds small until you picture two software requests that both count as one message: one might be a 20-line bug fix, and the other might be a full codebase review with thousands of lines. Token pricing separates those jobs the way a taxi meter separates a one-mile ride from a trip across town. (help.openai.com) OpenAI says the new Codex pricing already applies to new and existing ChatGPT Business plans and to new ChatGPT Enterprise plans. Existing ChatGPT Plus, ChatGPT Pro, and Enterprise or Education customers stay on the legacy rate card for now, but OpenAI says it will migrate them “in the upcoming weeks.” (help.openai.com) The company repeated the same split in separate release notes: new Enterprise workspaces use token-based rates, while existing Enterprise and current Education, Teachers, and Healthcare workspaces remain on message-based pricing until migration. ChatGPT Business release notes also warn that the switch “may result in a difference” in what customers are charged. (help.openai.com 1) (help.openai.com 2) OpenAI is also reshaping packaging around the product, not just the meter. Its Enterprise help page says it is introducing a “Codex seat,” a seat type with Codex-only access that uses flexible pricing instead of the standard all-in workspace bundle. (help.openai.com) This is becoming the normal pattern in artificial intelligence software. Anthropic’s official Claude pricing page says its current structure applies to Claude Sonnet 4.5, Claude Haiku 4.5, and future models, while earlier models keep older prices, so even within one vendor the bill can depend on which generation you are using. (platform.claude.com) Anthropic also prices extra behavior separately: the same page lists different rates for base input tokens, output tokens, and cache writes, which are stored prompts reused later to save work. That means one customer’s cost can change not just with usage volume, but with product design choices like how often an app reuses context. (platform.claude.com) For procurement teams, this is the hard part. A seat price is like a gym membership you can budget in January, while token pricing is like a cloud-computing bill that moves with every heavy user, long prompt, and automated workflow. (help.openai.com) (platform.claude.com) That is why a pricing-page update matters beyond Codex itself. The big artificial intelligence vendors are still deciding whether coding tools should be sold like software licenses, like cloud infrastructure, or as a hybrid of both, and customers are being moved between those models in real time. (help.openai.com) (platform.claude.com)
