Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Move Cursor work before the OpenAI shutoff

Written November 2, 2026

OpenAI proposed ending model supply to Cursor on November 12. Check what you run on GPT models, test the replacements, and know where BYOK does not reach.

What is proposed

OpenAI says it is planning to wind down its contract providing OpenAI models to Cursor, and that it has proposed a transition period of November 12, 2026. During that period, OpenAI says it would continue providing Cursor access to the OpenAI models it uses today. That date is a proposal, not a confirmed end date: the same page says Cursor may choose to end access sooner, and that OpenAI will share the official termination date as soon as it is confirmed between both companies. No entry about the shutoff appears on Cursor's changelog page. Work from the proposed date, and re-check the OpenAI page before treating it as settled.

Before you start

List every place your work depends on an OpenAI model inside Cursor: the model you pick in Chat and Agent, and any work that runs through Tab, Auto, Cloud or Background Agents, Automations, the Cursor CLI, or Cursor's API and SDK. The distinction matters, because a personal API key rescues only some of those surfaces. Note which prompts and tasks each one carries, so the replacement test in step 2 covers real work, not a greeting.

Steps

  1. Inventory what runs on GPT models. Open the model picker in Chat and Agent and write down each OpenAI model you use and what it does for you. Then list the unattended surfaces separately — Tab, Auto, Cloud or Background Agents, Automations, the CLI, and the API and SDK — because the key and gateway routes below do not reach them.
  2. Test the replacements on the same work. Pick the models available in your own Cursor model picker that you would use instead, and run the same prompts and tasks from your inventory on each one. Compare the results against the GPT-model results you already have. A replacement that has not run your actual task is not a replacement yet. If a task depends on an OpenAI model with no workable replacement in the picker, that task is the one the key, extension, and gateway routes in the next steps must cover.
  3. Use your own OpenAI API key for local Chat and Agent, if that fits. OpenAI's page describes this route: create an OpenAI API key on an account with billing configured, open Cursor Settings and select Models, paste the key, save or verify it, enable the model you want, and select it in Chat or Agent. Requests are billed through your OpenAI API account, separately from your Cursor or ChatGPT subscription, and a ChatGPT subscription does not include API usage. Cursor's own-key documentation adds that custom keys work only with chat models, that Tab completion keeps using Cursor's built-in models, and that Cursor's Zero Data Retention policy does not apply when you use your own keys.
  4. Or run Codex directly inside Cursor. The Codex IDE extension runs as a separate extension and does not depend on OpenAI models being available through Cursor's own model picker. Sign in with an eligible ChatGPT subscription or use an API key, and work from the Codex panel. It does not change the model used by Cursor Chat, Agent, Tab, Auto, Cloud Agents, or other Cursor services — it is a separate agent experience inside the editor.
  5. Or connect through a gateway you already use. OpenAI's page describes connecting Cursor to OpenAI models through an account with a compatible provider such as Amazon Bedrock or Azure, or another OpenAI-compatible gateway, using the provider integration or a custom base URL. Gateway compatibility varies, and some model-specific controls may not be available through a custom base URL. Gateway credentials, like a personal key, cover only supported local Chat and Agent requests.
  6. Re-check the date before November 12. Re-read OpenAI's page and Cursor's changelog. If the two companies confirm a different official date, or Cursor ends access sooner, the inventory from step 1 tells you exactly which work moves and which surfaces a key cannot rescue.

Where a personal key does not reach

OpenAI's page is explicit: Cursor does not support "Bring your own key" in Cursor Tab, Auto, Cloud or Background Agents, Automations, the Cursor CLI, or Cursor's API and SDK. Those features use models supplied or routed by Cursor and cannot use your OpenAI API key or gateway credentials. The key route also does not apply to Cursor Tab and autocomplete, Auto model selection and routing, or Cursor-provided or proprietary models. Work on those surfaces needs a replacement model from Cursor's own supply, tested in step 2 — not a key.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Why Cursor ignores your rules, and how to fix it · Cursor starter kit · Pick the cheapest model for the job