OpenAI is winding down the contract that supplies GPT models to Cursor, with a proposed shutoff of 12 November 2026. Two things matter more than the date. First, 12 November is a ceiling, not a deadline — OpenAI is cancelling at the latest date its change-of-control clause allows, and Cursor may cut access sooner. Second, bring-your-own-key is not a workaround: a personal key restores local Chat and Agent and does not restore Tab, Auto mode, cloud agents, Automations or the CLI. Our picks: staying on Cursor and switching models is free and fixes most cases; if you need GPT specifically, Codex at $20; if you need breadth, Copilot Pro at $10.
Why the date is the least useful fact here
Almost every write-up of this leads with 12 November 2026. That is the wrong anchor, in a way that costs you if you plan around it.
OpenAI announced on 28 August 2026 that it had notified SpaceX of its intent to “wind down our contract providing OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026,” two weeks after SpaceX closed its $60 billion acquisition of Anysphere on 14 August. The stated reason is counterparty conduct rather than competition: OpenAI said it cannot be confident SpaceX will use its technology within OpenAI’s terms of service, citing prior experience with Musk-owned companies violating contracts.
Three qualifiers in that announcement get dropped in summary, and together they invert how you should plan:
- The date is proposed, not agreed. OpenAI has said the official termination date will be published once both companies confirm it, and Cursor’s CEO has said discussions are continuing.
- 12 November is the latest date allowed, not a chosen one. OpenAI’s framing is that the change-of-control provision opens a limited cancellation window and it is cancelling at the last date that window permits. The date is an artefact of the contract’s outer bound.
- Cursor may end access sooner. OpenAI’s own guidance says so explicitly.
Put those together and the asymmetry is clear. There is a documented mechanism by which GPT access in Cursor disappears before 12 November, and no stated mechanism by which it survives past it absent a settlement. A migration scheduled for early November has slack in the wrong direction.
What a personal API key actually restores
This is the part the traffic-share headline buries. The question is not whether you can keep calling GPT — you can — but on which surfaces.
| Surface | Works with your own OpenAI key? |
|---|---|
| Local Chat (desktop app) | Yes |
| Local Agent (desktop app) | Yes |
| Tab / autocomplete | No — stays on Cursor’s built-in models |
| Auto mode (automatic model selection) | No |
| Cloud and background agents | No |
| Automations | No |
| Cursor CLI | No |
| Bugbot and agentic code review | No |
Cursor’s own documentation is the primary source for the first constraint: “custom API keys only work with chat models,” and “tab completion continues using Cursor’s built-in models.” OpenAI’s guidance for Cursor users supplies the rest of the list, confirming that a personal key covers local Chat and Agent but not Tab, Auto, Cloud Agents, Automations or the CLI.
The limitation is architectural, which is why it is unlikely to soften. Those surfaces execute on Cursor’s infrastructure, so the call reaches OpenAI from Cursor, as a customer — the exact commercial relationship being terminated. Your key authorises your client. It cannot re-admit Cursor’s servers as a permitted buyer.
The practical consequence is that this lands hardest on the tier Cursor spent 2026 pushing hardest. Cursor’s cloud agents and always-on subscriptions are the flagship feature, and they are on the “No” side of that table. A pinned gpt-5.x string inside an Automation or a CI job is not a preference you change in a dropdown on the day. It is a scheduled outage.
The markup BYOK does not escape
Here is the arithmetic no vendor page performs, and it is the reason “just use your own key” is worse advice for teams than for individuals.
Cursor bills third-party models from its “Other Models” pool at each model’s own API rate, so on an individual plan, paying OpenAI directly is roughly cost-neutral. On Teams and Enterprise it is not. Cursor’s billing documentation states that on those plans “third-party model requests include a Cursor Token Rate of $0.25 per million tokens,” and that this rate applies on top of model API pricing for included usage, on-demand usage and BYOK usage.
So a team that responds to the cutoff by having every developer paste in a personal OpenAI key keeps paying Cursor per token for the privilege:
| 20-seat team, Teams Standard | Monthly |
|---|---|
| Seats (20 × $40) | $800 |
| Cursor Token Rate on 300M third-party tokens | $75 |
| Your own OpenAI API invoice | at OpenAI list |
| Features lost on those tokens | Tab, Auto, cloud agents, Automations, CLI |
Cursor’s first-party models — Grok 4.7, 4.6, 4.5 and Composer 2.5 — are exempt from the Cursor Token Rate and sit in a separate pool with “significantly more included usage.” That is not a neutral rate card. It is a price signal, and after 12 November it is a louder one.
The exits, priced
Prices verified 5 October 2026 against each vendor’s pricing page: Cursor, Claude, GitHub Copilot, Devin.
| Option | Cost / dev / month | OpenAI models? | Best when |
|---|---|---|---|
| Stay on Cursor, switch model | $20 (Pro), unchanged | No (first-party pool) | You chose the editor, not the model |
| Codex | $20 (ChatGPT Plus), $200 (Pro) | Yes, first-party | You specifically need GPT behaviour |
| Codex CLI + API key | Free tool + API usage | Yes, at API rates | You want GPT without a subscription |
| GitHub Copilot Pro | $10, incl. $15 AI credits | Yes, alongside Claude/Gemini/xAI | You want breadth from a neutral vendor |
| Claude Code | $20 (Pro) → $100/$200 (Max) | No, Anthropic only | Terminal-first agentic work |
| Devin Desktop (ex-Windsurf) | $20 (Pro), $200 (Max) | Via Cognition’s roster | You want a Cursor-shaped editor elsewhere |
| OpenCode / Aider | Free + your API spend | Yes, BYOK any provider | You are comfortable in a terminal |
Three notes the table cannot carry.
Copilot Pro is the value answer for mixed-model work. At $10 with $15 of included credits it is the only option here that is cheaper than Cursor and keeps OpenAI, Anthropic, Google and xAI models on one plan — and GitHub’s relationship with OpenAI is not downstream of this dispute. Our Copilot vs Cursor comparison covers where it still trails on editor feel.
Check what Windsurf is before you migrate to it. Cognition renamed the editor Devin Desktop on 2 June 2026 and retired the Cascade agent on 1 July, replacing it with Devin Local; windsurf.com now redirects to devin.ai. It is a rename and consolidation rather than a shutdown, and existing installs updated in place — but if your shortlist says “Windsurf” because of a 2025 review, you are shortlisting a product that no longer exists under that name. See our Windsurf review and Cursor vs Windsurf.
Staying is the cheapest fix and the right one for most people. If you sit in Auto mode or deliberately pick Claude, Gemini or Grok, nothing in your workflow depends on the supply being terminated. Grok 4.7 ships at the same rate card as 4.6 and sits in the included pool. Zero cost, zero migration.
Migration checklist
- Grep for pinned model strings, not for the word “OpenAI.” Search your repo, CI config, Automations, cloud agent definitions and Cursor CLI invocations for
gpt-. A pinned ID in automation is the failure mode; a dropdown selection is not. - Sort the hits by surface, using the table above. Anything on the “No” side needs a model change before the cutoff. Anything in local Chat or Agent can fall back to a personal key.
- Do not schedule for November. Treat 12 November as the outer bound and target mid-October, because the date can only move earlier.
- If you are on Teams, price the Cursor Token Rate before choosing BYOK. $0.25 per million tokens applies to BYOK usage too; at scale, that can exceed the cost of simply switching models.
- Validate the replacement on your own repository. You presumably picked a specific GPT model because of behaviour you noticed on your code, so a week on real tickets beats a benchmark table.
- Re-check the date in early November. Talks are continuing, and a settlement turns this page from a migration plan into a vendor-risk post-mortem.
The buying lesson worth keeping
The reason to read this as more than a Cursor story is that the lever used here is generic. Model neutrality in a multi-model tool was never an architecture; it was a contract term, and contract terms have counterparties and change-of-control clauses. Cursor’s dropdown looked like optionality and was in fact four commercial agreements, one of which has now been cancelled by the supplier over who bought the customer.
What makes it structural rather than a one-off is the forward half: future OpenAI models are reportedly not to be supplied to Cursor at all. That converts a fixed migration into a widening gap, and it is the part to price at renewal. For the wider field, see our best AI coding tools and best AI harnesses; for the head-to-head most readers land on next, Cursor vs Claude Code. Our original analysis of the termination is here.
What changed
- 5 Oct 2026: Page created. 12 November re-verified as still proposed and unconfirmed, with talks continuing — and clarified as the latest date OpenAI’s change-of-control window allows, not a chosen deadline.
- 5 Oct 2026: Confirmed from Cursor’s billing docs that the $0.25/M Cursor Token Rate on Teams and Enterprise applies to BYOK usage, so a personal key does not escape Cursor’s per-token charge.
- 2 Jun – 1 Jul 2026: Windsurf renamed Devin Desktop; Cascade retired 1 July. Relevant because it is widely listed as a Cursor exit under a name it no longer uses.
- 30 Aug 2026: OpenAI moved to end model supply to Cursor, proposing 12 November 2026 and no future OpenAI models at all.
- 14 Aug 2026: SpaceX’s $60B acquisition of Anysphere closed, opening the change-of-control window OpenAI later exercised.
Frequently asked questions
Will OpenAI models really stop working in Cursor on 12 November 2026?
Probably, but the date is the wrong thing to plan against, and not for the reason most coverage gives. OpenAI's statement was that it intends to wind down the contract "with a proposed shutoff date of November 12, 2026" — proposed, not agreed. The official date is to be published once both companies confirm it, and Cursor's CEO has said talks are continuing. What almost no coverage makes explicit is which direction the date can move. OpenAI has said it is exercising the cancellation window its change-of-control clause opens and cancelling at the latest date that provision allows, which is what produces 12 November; it has also said Cursor may choose to end access sooner. Both halves point the same way: 12 November is a ceiling, not a target. It can arrive earlier if Cursor cuts over first or the two sides settle on something nearer, and the contractual mechanism does not appear to allow it to drift later. If your plan is to migrate during the first week of November, your plan has no slack in the only direction that matters.
Can I just paste in my own OpenAI API key?
For interactive editing, largely yes. For anything automated, no — and this is the gap the headline traffic-share number hides. Cursor's own documentation states that "custom API keys only work with chat models" and that "tab completion continues using Cursor's built-in models." OpenAI's guidance for Cursor users goes further, confirming that a personal key covers local Chat and Agent but does not cover Tab, Auto mode, Cloud Agents, Automations or the Cursor CLI. The reason is structural rather than a policy that might soften: those surfaces run on Cursor's own infrastructure, so the request reaches OpenAI from Cursor as a commercial customer — which is exactly the relationship being terminated. A key you hold can only authorise the paths where your own client talks to OpenAI directly. It cannot re-admit Cursor's servers. Treat bring-your-own-key as a fallback for typing in the editor, not as a continuity plan for a pipeline.
Does bringing my own key at least save money?
On an individual plan it is roughly cost-neutral, and on Teams or Enterprise it is explicitly not. Cursor bills third-party models from its "Other Models" pool at the model's own API rate, so paying OpenAI directly instead lands in much the same place. But Cursor's billing documentation adds that on Teams and Enterprise plans "third-party model requests include a Cursor Token Rate of $0.25 per million tokens," and states that this rate applies on top of model API pricing for included usage, on-demand usage, and BYOK usage alike. Read that last clause carefully: supplying your own OpenAI key does not exempt you from Cursor's per-token charge. A twenty-seat team pushing 300 million third-party tokens a month pays roughly $75 in Cursor Token Rate on top of its own OpenAI invoice, on top of $800 in seats — for a configuration that has lost Tab, Auto and cloud agents on those models. Cursor's first-party models, Grok and Composer, are exempt from the rate, which tells you where the pricing is steering you.
Is the "5% of traffic" figure a reason to relax?
It is accurate and it is measuring the wrong quantity. Cursor's CEO put OpenAI models at roughly 5% of user traffic, and as a share of tokens that is almost certainly right — most Cursor work runs through Auto mode, Claude, Gemini, Grok or Cursor's own Composer, none of which are affected. The problem is that traffic share is a volume measure and migration risk is a dependency measure. A single pinned GPT model string inside a cloud agent, an Automation or a CLI job in CI is a rounding error in token share and a total outage in that workflow, because those are precisely the surfaces a personal API key cannot rescue. The honest reading is that this affects few users a lot rather than many users a little. Find out which group you are in by auditing pinned model IDs, not by looking at your token dashboard.
What should I actually switch to, and what does it cost?
If your reason for using GPT in Cursor was the model, move the work to a tool that still has a first-party relationship with OpenAI: Codex is included with ChatGPT Plus at $20 a month and Pro at $200, and the Codex CLI is free and open source if you would rather pay OpenAI API rates directly. If your reason was the editor rather than the model, staying on Cursor Pro at $20 and moving to Claude, Gemini, Grok 4.7 or Composer 2.5 is the cheapest fix available, because it costs nothing and changes one dropdown. If you want model breadth under a vendor that is not downstream of this dispute, GitHub Copilot Pro is $10 a month including $15 of AI credits and carries OpenAI, Anthropic, Google and xAI models on one plan — the best value on this list for mixed-model work. Claude Code is included in every paid Claude plan from $20 and is the strongest terminal agent, but it is Anthropic-only by design. Avoid treating Windsurf as an exit without checking what it is now: Cognition renamed it Devin Desktop on 2 June 2026, retired the Cascade agent on 1 July, and windsurf.com now redirects to devin.ai.
Should this change whether we renew Cursor at all?
Not on its own, and yes if you were already close to the line. Losing one of four model suppliers is a real reduction in a product whose central pitch is that you get every model, but Cursor keeps Anthropic, Google and xAI, plus its own Composer line, and for most users the practical capability change is close to zero. The thing that deserves weight at renewal is not the loss itself but its shape. OpenAI's position is reported as covering future models too — Cursor's OpenAI shelf is frozen at today's contents and then removed — so this is a widening gap rather than a one-time migration, and every subsequent OpenAI release is a capability competitors can offer and Cursor cannot. Combine that with ownership: under SpaceX, Cursor has clear incentives to route work toward Grok, and its pricing already gives first-party models a dedicated pool with "significantly more included usage" and exemption from the Teams token rate. If you bought a neutral aggregator, price your renewal against what the product is becoming rather than what it was.