Cursor OpenAI Models: November 12, BYOK, Codex and Gateways

To keep using OpenAI models in Cursor, bring your own API key (BYOK) for supported local Chat and Agent requests. Choose the Codex IDE extension for its coding agent or access through an eligible ChatGPT subscription. Use a compatible gateway if your team already manages model access and billing there.

These routes have different account requirements and feature limits. Check those before buying API credit or changing your endpoint.

November 12 remains a proposed transition date

As of 2026-10-27, November 12, 2026 is still a proposed termination date. OpenAI’s current Cursor guidance says the final date awaits confirmation between both companies. It also says Cursor could end access sooner.

The timeline has two relevant stages:

2026-08-28: OpenAI announced its intended contract wind-down following SpaceX’s acquisition of Cursor. It proposed November 12 and said it would withhold future models from the agreement.
As of 2026-10-27: OpenAI’s help page still describes that date as proposed. It documents BYOK, the Codex extension, and compatible gateways as alternative access routes.

OpenAI attributed its decision to concerns about compliance with its terms. That is OpenAI’s stated rationale. For migration planning, the practical issue is the supply agreement behind Cursor’s own OpenAI access.

Configure an alternative before relying on the proposed deadline. These instructions preserve supported ways to work inside the editor; they do not preserve every Cursor feature.

Check your plan, feature, and bill first

Cursor’s native BYOK facility covers supported local Chat and Agent requests. It does not cover Tab, Auto, Cloud or Background Agents, Automations, Cursor CLI, or Cursor API/SDK. Installing the Codex extension does not change those Cursor features.

Plan eligibility also matters. As of 2026-10-27, Cursor’s current plan documentation explicitly excludes personal API keys from Cursor Plus. Its BYOK help covers Premium, Super, Ultra, and specified legacy plans. Check your existing entitlement before changing plans.

On an Enterprise workspace, an administrator can restrict personal keys. The controls are under Team Settings → Models → Personal API Keys (BYOK). A gateway entered through the OpenAI fields needs permission for OpenAI BYOK too.

The payment routes are separate:

Route	What pays for model usage
OpenAI key in Cursor	Your OpenAI API account
Codex with ChatGPT sign-in	Your eligible ChatGPT plan’s allowance and applicable additional credits
Codex with an OpenAI API key	Your OpenAI API account
Gateway in Cursor	Your gateway or hosting-provider account

API charges are separate from Cursor and ChatGPT subscription fees. A ChatGPT subscription does not provide an OpenAI API balance.

Teams and Enterprise have another cost to check. As of 2026-10-27, eligible BYOK requests carry a Cursor Token Rate of USD $0.25 per million tokens. It applies to input, output, and cached tokens, in addition to provider charges. Individual plans do not pay this rate, and listed model exemptions also apply to BYOK.

“Local” describes the desktop workflow. According to Cursor’s BYOK documentation, requests still pass through Cursor’s backend for prompt construction. Cursor’s Zero Data Retention policy does not apply to BYOK. Check your provider’s data handling policy.

Route 1: Add your own OpenAI API key

Choose this route when you want Cursor’s native Chat and Agent interface with a supported OpenAI model.

Cursor currently documents OpenAI BYOK for standard chat models without reasoning. Do not assume that adding a newer reasoning-model ID makes it supported. Choose from the models your current client and API account support.

Follow the OpenAI setup instructions for Cursor:

Configure billing for your OpenAI API account.
Create your own key on the OpenAI API keys page.
Open Cursor Settings → Models and find the OpenAI API key settings.
If you previously configured a gateway, turn off Override OpenAI Base URL before using your official OpenAI key.
Paste the key into OpenAI API Key and click Save. Enable Use OpenAI API Key if your interface shows that switch.
Enable a supported OpenAI model in the model list.
Open a new local Chat, turn off Auto, and explicitly select that model.
Send Reply with exactly: connection-ok.

The expected result is a reply in Cursor Chat. Check the corresponding project’s API usage to confirm which account received the request. Then test the Agent operation you actually need, using the small editing check below.

Route 2: Install the Codex IDE extension

Choose Codex when you need its coding-agent workflow or want to use eligible ChatGPT subscription access inside Cursor. It runs in its own interface and does not depend on Cursor’s native OpenAI model picker.

Eligible ChatGPT plans include Plus, Pro, Business, Edu, and Enterprise, subject to workspace permissions. Check current Codex plan availability. ChatGPT Plus and Cursor Plus are different subscriptions.

Open the official Codex IDE extension page.
Under Install or enable Codex, choose Cursor and complete the installation.
Open a local project, then select the Codex icon in the Activity Bar.
If the icon is missing, open the Command Palette and run Codex: Open Codex Sidebar.
Choose Sign in with ChatGPT and complete the browser sign-in. For API billing, choose Use API Key, enter your OpenAI key, and select OK.
Open a source file and add a selected function to the Codex prompt context.
Ask Codex to explain the function and one edge case without editing files.

Read the answer in the Codex conversation and compare it with the selected code. Use Codex’s own model and account controls when checking the connection.

For agent settings, select the sidebar’s gear icon, then Codex Settings. OpenAI’s extension settings reference separates these controls from editor settings. They do not configure Cursor’s native Models page.

The authentication documentation confirms that API-key sign-in bills your OpenAI Platform account. It supports local Codex work, while features requiring ChatGPT workspace or cloud access can remain unavailable.

Submit future tasks in the Codex panel if you choose this route. Typing into Cursor Chat still uses Cursor’s separate configuration.

Route 3: Connect an OpenAI-compatible gateway

Choose a gateway when it provides the account, model, and API format your Cursor workflow needs. OpenAI lists Amazon Bedrock, Azure, and other compatible providers as possible routes.

Azure and Bedrock have provider-specific settings. The steps below cover a custom OpenAI-compatible endpoint; use the provider’s deployment instructions for its credentials and model names.

For OpenRouter-specific steps, see the separate Cursor–OpenRouter setup guide.

Obtain your provider’s API key, documented Base URL, and exact model ID.
Confirm that the provider supports the API format and tools required by your selected Cursor model.
Open Cursor Settings → Models → API Keys.
In the OpenAI section, enable Override OpenAI Base URL and enter the provider’s Base URL.
Enter the gateway’s credential in OpenAI API Key, then enable Use OpenAI API Key and save.
Add or enable the exact model ID supplied by the provider.
Restart Cursor, open a new local Chat, disable Auto, and select that model.
Send the short connection request, then inspect the provider’s usage record.

Cursor support documents these controls in its custom-model setup guidance. Use the endpoint’s model ID, including any required prefix. A display label from another provider may identify a different route.

For example, BetterToken’s Cursor guide specifies https://www.bettertoken.ai/v1 and your own BetterToken API key. Its current catalogue listed gpt-6-luna on 2026-10-27. That confirms catalogue availability, not native Cursor compatibility. Check the model’s API mode and tool support before selecting it, especially given Cursor’s published reasoning-model restriction.

For the BetterToken Base URL field, retain /v1. Do not append /chat/completions or /responses; those are resource paths, as its API reference explains.

Check the override when switching back to Cursor-managed models. In Cursor’s routing support thread, support confirmed that the override also affects built-in OpenAI model selections. Disable it when returning to the Cursor-managed route. Choosing a familiar GPT label alone does not select a different endpoint.

Verify the full editing path

Test the operation you need before moving a coding session to the new route. A short reply establishes basic connectivity, but it does not prove that editing tools work.

Open a disposable project folder. In Cursor’s local Agent or the Codex sidebar, submit:

Create cursor-route-check.txt containing this single line:
connection-ok

Use your file-editing tool, then read the file back.
Do not use a shell command to create or modify the file.
Stop after confirming its contents.

The expected result is a file containing connection-ok and a completed edit-and-read-back interaction. Inspect the file yourself. For API-key or gateway access, also check the provider’s recorded model and usage.

If text works but the edit fails, investigate tool compatibility before replacing the key. In a July 2026 Cursor support thread, Harm Alhusen traced failed edits to his proxy dropping custom_tool_call responses. Cursor support explained that OpenAI custom tools must survive the full request-and-response path.

That report describes one diagnosed integration failure. It does not establish the cause of every aborted Agent task. If your required model cannot complete the test through native BYOK, try the Codex route with a supported account and model.

Fix the matching failure

Use the symptom to choose the next check. Changing every setting at once makes routing errors harder to isolate.

Symptom	Check and next action
API-key or Base URL fields are missing	Check your Cursor plan and client. On Enterprise, ask the administrator to review personal-key restrictions. Reopen Models after access is enabled.
Authentication fails	Confirm that the key belongs to the configured endpoint’s provider and that API billing is active. Save the corrected pair, then retry a short request.
`AI Model Not Found` or `Model name is not valid`	Check the exact provider model ID, the active key switch, and OpenAI BYOK permission. Start a fresh chat after correcting them.
`You've reached your monthly usage limit` while using a gateway	Check Cursor team spending limits as well as provider credit. Applicable Cursor token charges can still stop team BYOK requests.
A built-in GPT selection fails after enabling an override	Disable the override before testing the Cursor-managed route, while that model remains available there.
A reply starts, but tools or edits fail	Check protocol and tool-call handling for that specific model. Repeat the disposable-file test after the provider or client configuration is corrected.

The September–October 2026 Cursor support exchange documents several of these permission and spending-limit failures. They are useful checks, not proof that your account has the same cause.

If the request still fails, record the Cursor version, selected model ID, endpoint host, timestamp, and request ID. Keep credentials and private prompt content out of shared diagnostics.