OpenAI rolls out GPT tool actions and Codex pricing

OpenAI has started turning two quiet product details into policy: which models a custom assistant can call, and how much a coding agent costs once people actually use it. In recent Help Center updates, the company said GPTs with Custom Actions can use GPT-4o and GPT-4.1 on the web for ChatGPT Plus, Pro, and Team users, with Enterprise and Edu “coming soon,” and it separately published a Codex rate card that maps credit usage across plans. (help.openai.com 1) (help.openai.com 2) That may sound like housekeeping, but it changes how OpenAI packages automation. A custom GPT is the user-facing assistant, while a Custom Action is the connector that lets that assistant reach outside ChatGPT to call an external application programming interface, which is a software bridge between services. Once those actions are tied to named model options and plan tiers, “build a GPT” stops being just a creative feature and starts looking more like a managed software product. (help.openai.com 1) (help.openai.com 2) The release-notes change matters because model choice affects how a GPT behaves when it uses tools. GPT-4.1 and GPT-4o are different model families with different performance profiles, so letting builders pick them inside Custom Actions gives paying users more control over speed, reasoning style, and tool orchestration on the web version of ChatGPT. OpenAI’s note places that capability in Plus, Pro, and Team first, then says Enterprise and Edu are next. (help.openai.com) There is also a timing wrinkle. OpenAI’s newer Help Center pages say GPT-4o and GPT-4.1 were retired from ChatGPT on February 13, 2026, with GPT-4o lingering inside some business custom-GPT contexts until April 3, 2026, before full retirement across plans. That means the release-note language cited in this story appears to reflect a transitional moment in OpenAI’s rollout, rather than the final steady state users see today on April 8, 2026. (help.openai.com 1) (help.openai.com 2) (help.openai.com 3) The second update is less visible to casual users and more important to teams that want coding agents in production. OpenAI’s Codex rate card says that, as of April 2, 2026, Codex pricing for new ChatGPT Business and new ChatGPT Enterprise plans shifted to align with application programming interface token usage instead of per-message pricing. Existing Plus, Pro, and Enterprise or Edu customers remain on a legacy rate card until migration in the coming weeks. (help.openai.com) That pricing change does more than alter a bill. Per-message pricing is easy to understand but blunt, because one short request and one long multi-file coding task can count the same way. Token-based pricing meters the amount of text and model work more directly, which makes costs track usage more precisely when Codex is reviewing code, generating patches, or operating across bigger repositories. (help.openai.com) (help.openai.com) OpenAI is also making credits the common language across products. Its credits documentation says the same flexible credits can be used across supported features, and today that includes Codex for Plus and Pro subscribers and Sora video generation for eligible users. In practice, that turns credits into a shared wallet for high-compute features rather than a product-specific coupon. (help.openai.com) For companies, the credit system is becoming part of procurement, not just usage. OpenAI’s business documentation says Codex seats require credits for activity, and workspace administrators can add credits, set auto-top-ups, and monitor usage analytics and spend controls. That means deploying an internal coding assistant now involves finance settings and governance rules alongside model selection. (help.openai.com) (help.openai.com) The combination of Custom Actions and Codex pricing points to the same product direction. OpenAI is moving from “chat with an assistant” toward “assign work to an assistant that can call tools, touch outside systems, and consume a metered budget.” One update governs what the assistant is allowed to connect to; the other governs what happens when those connections start doing real work at scale. (help.openai.com) (help.openai.com) That matters most for two groups. Individual power users on Plus and Pro get earlier access to more capable automation inside custom GPTs and a clearer path to paying for heavier coding workloads. Businesses get a more explicit operating model: choose the plan, choose the models, turn on actions, buy credits, and watch spend. (help.openai.com) (help.openai.com) (help.openai.com) There is still some ambiguity because OpenAI’s Help Center now contains overlapping notes from different rollout stages. The strongest current signal is that the company is standardizing around newer model lineups in ChatGPT while preserving temporary legacy access in certain business contexts and shifting Codex toward token-based metering for more plans over time. That makes the story less about one feature launch than about OpenAI hardening the commercial rules around agents. (help.openai.com) (help.openai.com) (help.openai.com) If this pattern holds, the next phase of competition in artificial intelligence assistants will not just be about model quality. It will be about who gives developers and administrators the cleanest way to connect tools, predict cost, and decide which users get which level of autonomy. OpenAI’s latest notes show those controls moving out of the background and into the product itself. (help.openai.com) (help.openai.com)