TL;DR: On August 28, 2026, OpenAI announced it will end Cursor’s direct access to its models on November 12, invoking a change-of-control clause after SpaceX closed its $60 billion acquisition of Anysphere on August 14. If you run Composer, Claude, Gemini, or Grok inside Cursor — Cursor CEO Michael Truell says that’s about 95% of user traffic — nothing breaks. If you depend on the GPT-5.6 family in Cursor’s model picker, you have 74 days. Your own OpenAI API key is the obvious fallback, and OpenAI itself points to it, but Cursor has not confirmed how in-app key support behaves after the cutoff — so test your fallback now, not on November 11. Here’s the full picture, verified August 30, 2026.

What OpenAI actually announced

The timeline is unusually compressed for a deal this size. SpaceX finalized its $60 billion all-stock acquisition of Anysphere, Cursor’s parent company, on August 14, 2026. Exactly two weeks later, on August 28, OpenAI published a post titled “Our decision on Cursor following its acquisition by SpaceX” and notified SpaceX that it intends to wind down the contract that supplies OpenAI models to Cursor. Proposed shutoff: November 12, 2026, which OpenAI describes as “the maximum notice our contract allows.”

The stated reason is not capacity or economics. OpenAI says it invoked a change-of-control clause because it cannot be confident SpaceX will use its technology within OpenAI’s terms of service, citing what it characterizes as a pattern at Elon Musk’s companies — X breaching contract terms after Musk acquired it, and xAI, which OpenAI says Musk admitted under oath had violated OpenAI’s terms.

Cursor’s response came the next evening. Truell posted on X that “OpenAI models serve about 5% of Cursor user traffic, and we’re speaking with the OpenAI team to resolve this.” That 5% figure is the single most important number in this story: for nineteen out of twenty Cursor users, November 12 is a headline, not an outage.

One more move worth noting: within hours of OpenAI’s announcement, Anthropic co-founder Tom Brown confirmed Anthropic will increase compute capacity for Claude models on Cursor. Google has said nothing publicly, but Gemini models remain in the picker. Model supply just became a competitive weapon, and Cursor’s other suppliers are leaning in, not out.

What actually breaks on November 12

Cursor’s model lineup in late August 2026 looks like this:

Model family	Supplier	Status after Nov 12
Composer 2.5 (default)	Cursor in-house	Unaffected
GPT-5.6 Sol / Terra / Luna	OpenAI	Direct access ends
Claude Sonnet 5 / Opus 5	Anthropic	Unaffected — compute being increased
Gemini 3.x	Google	Unaffected
Grok 4.5 / 4.6	SpaceXAI	Unaffected (same parent company now)

The models that disappear from Cursor’s built-in credit pool are the GPT-5.6 family — Sol (the flagship, which runs in Max Mode only), Terra, and Luna. Everything else stays.

It’s worth being precise about what “direct access” means, because this is where the queue of hot takes gets sloppy. OpenAI is terminating the commercial contract under which Cursor buys inference from OpenAI and resells it through Cursor’s usage credits. That’s the path your Pro plan’s included usage rides on. Three things are not covered by that termination, per OpenAI’s own announcement:

1. Your own OpenAI API key. OpenAI explicitly points developers to personal API keys as a continuing path.
2. The Codex IDE extension, which OpenAI ships for VS Code and VS Code forks — Cursor included.
3. Compatible gateways that sit between an editor and the OpenAI API.

The catch — and it’s a real one — is that all three of those paths depend on Cursor’s side continuing to work smoothly, and Cursor has published no statement about migration support, credit handling, or whether in-app custom-key support for OpenAI models will behave identically after November 12. Truell’s post says the companies are still talking. A negotiated resolution is possible. Planning around one is not a plan.

The BYOK question, answered honestly

Cursor has supported custom API keys for years: Cursor Settings → Models → API Keys, paste your key, toggle it on. Here is what that actually gets you today, and where people get surprised.

What routes through your key: chat, Plan mode, and Agent mode. When a custom key is active, those requests bill to your OpenAI account at public API rates instead of drawing down Cursor credits.

What never routes through your key: Tab autocomplete and inline edit (Ctrl/Cmd+K). Both are locked to Cursor’s own backend — and that matters here in a reassuring way, because Cursor Tab runs on Cursor’s in-house model, not on OpenAI. Your autocomplete does not break on November 12 no matter what happens to the OpenAI relationship.

The problem you’ll actually hit: developers wiring up custom keys with a gateway or custom base URL regularly find that requests fail even with a valid key, because Cursor doesn’t send plain Chat Completions calls — OpenAI-bound requests use the Responses API format, and Anthropic-bound requests use a hybrid format that standard proxies mishandle. The fix is to skip the custom base URL entirely unless your gateway explicitly documents Cursor support: paste the bare key into the OpenAI API key field, click Verify, and confirm a live request in your OpenAI usage dashboard before you rely on it. If Verify fails with a valid key, check whether a team/enterprise admin policy or a privacy-mode setting is blocking custom keys — both are documented causes on Cursor’s forum.

Sanity-check the key itself outside Cursor first, so you’re debugging one thing at a time:

$ curl -s https://api.openai.com/v1/models \
    -H "Authorization: Bearer $OPENAI_API_KEY" | head -20
{
  "object": "list",
  "data": [
    {
      "id": "gpt-5.6-terra",
      "object": "model",
      "created": 1753833600,
      "owned_by": "system"
    },
...

If that returns a model list, your key and billing are live and any remaining failure is on the editor side.

The honest caveat: OpenAI saying “developers can keep using their own keys” is not the same as Cursor guaranteeing the in-app OpenAI key path works unchanged after the cutoff. Nobody outside the two companies can promise that today, and coverage that flatly claims “BYOK users are insulated” is asserting something neither company has said. That’s why the migration plan below has a plan B that doesn’t depend on Cursor’s OpenAI integration at all.

What BYOK costs vs. what you pay now

Cursor Pro is $20/month and includes $20 of model usage; Pro+ is $60 with $70 of usage; Ultra is $200 with $400. Against that, here are the public API rates you’d pay through your own key, verified August 30, 2026:

Model	Input $/M tokens	Output $/M tokens	Notes
GPT-5.6 Sol	$5.00 ($4.00 promo)	$30.00 ($20.00 promo)	Promo runs through at least Nov 21
GPT-5.6 Terra	$2.00	$12.00	Post-July-30 price cut
GPT-5.6 Luna	$0.20	$1.20	Fastest, cheapest tier
Claude Sonnet 5	$2.00	$10.00	Intro price now permanent
Claude Opus 5	$5.00	$25.00
Claude Haiku 4.5	$1.00	$5.00
Composer 2.5 (via credits)	$0.50	$2.50	Cursor in-house, no key needed

A worked example, so the table means something. Say your agent workload burns 1M input tokens and 100K output tokens a day (a moderate agent-mode day; heavy multi-agent users go far past this):

• GPT-5.6 Terra via your key: $2.00 + $1.20 = $3.20/day, roughly $70/month on workdays
• Claude Sonnet 5 via your key: $2.00 + $1.00 = $3.00/day — and it’s a lateral quality move for most coding work
• Composer 2.5 on Cursor credits: $0.50 + $0.25 = $0.75/day, comfortably inside Pro’s included $20/month

Two conclusions fall out of the math. First, if the GPT-5.6 cutoff pushes you to BYOK, Claude Sonnet 5 at $2/$10 is priced within a rounding error of Terra at $2/$12 — the “switch models” path costs the same as the “keep GPT via key” path. Second, Composer 2.5 is roughly 4x cheaper than either, which is exactly why Cursor can shrug off losing a 5%-traffic supplier. Prompt caching changes the absolute numbers on all of these (cached input is 10% of base on both providers), but not the relative ranking.

The migration plan, by user type

If you run Composer, Claude, Gemini, or Grok (the 95%): do nothing today. Your models, your Tab completion, and your pricing are untouched. The one thing worth doing this week is a five-minute insurance step: add one external API key (Anthropic or OpenAI) in Settings → Models, verify it, toggle it back off. You now have a tested fallback if any future supplier dispute hits a model you do care about. This event is the proof that single-supplier dependence inside a closed tool is a real risk, not a theoretical one.

If GPT-5.6 is your daily driver in Cursor: run the plan in this order, before November.

1. This week: create or fund an OpenAI API key, verify it works with the curl check above, add it in Cursor, and run a real agent session through it. You want the failure modes surfaced while the built-in path still works.
2. Also this week: install the Codex IDE extension inside Cursor as plan B. It’s OpenAI’s own extension, it runs in VS Code forks, it authenticates against your ChatGPT plan or API key, and it does not depend on the Cursor–OpenAI contract at all. This is the path OpenAI itself names, and it survives any outcome of the negotiation.
3. Before November 12: pick your steady state. Realistic options: Terra via your own key (same workflow, new bill), Claude Sonnet 5 in Cursor (near-identical cost, no key drama, supplier actively adding compute), or Composer 2.5 (cheapest, and the default for a reason — see our Cursor 3 review for how it holds up on agent work).
4. If the whole episode sours you on model-supplier politics: the tools that never had this problem are the ones where you always brought your own key or your own weights. Cline is BYOK-native. Note that Continue.dev is no longer the escape hatch it used to be — Cursor acquired it and froze the extension at v2.0.0 in July (what still works). Devin Desktop (the former Windsurf) is model-agnostic by design (our rebrand breakdown). For open-source agents with zero vendor lock-in, aifoss.dev tracks the FOSS side of this space.
5. The nuclear option — own the weights: a local model can’t be turned off by a press release. Qwen3.8-27B runs as a genuine Cursor-quality backend on a 24GB card (setup guide), and our Cursor + local Llama hardware tiers piece maps what $500–$3,000 of hardware buys. For the GPU side of that decision, runaihome.com covers local AI hardware in depth.

If you’re on Cursor Teams: your admin controls whether custom keys are allowed, and usage pools work differently (Teams pricing breakdown). Raise the November 12 date with whoever owns the seat contract now — enterprise procurement moves slower than 74 days.

Honest take

OpenAI just demonstrated that a model vendor will fire a customer’s customers to make a point in a corporate feud, with 76 days’ notice, and frame it as generosity. The 5% traffic figure means the blast radius is small this time. The precedent is not small. Every closed tool that bundles a single-provider model relationship — which is most of them — now carries visible counterparty risk that BYOK-native and local-first setups simply don’t have.

For Cursor users, the practical verdict: November 12 is not the end of Cursor, and for most people it’s not even an inconvenience. But “my editor’s model lineup depends on whether two CEOs are feuding” is a sentence you should only have to think once before setting up a fallback. Add the key, test it, and go back to work.

FAQ

Does my Cursor Tab autocomplete stop working on November 12? No. Tab runs on Cursor’s in-house model, not OpenAI, and never routed through the OpenAI contract.

Will my own OpenAI API key definitely keep working inside Cursor? OpenAI says personal keys and its Codex extension remain available. Cursor has not published a statement confirming how in-app key support behaves after the cutoff. Probable: yes. Guaranteed: no. Test it now, and keep the Codex extension installed as a fallback that bypasses the question entirely.

Do Cursor Pro credits get refunded for the GPT models I can no longer use? Unknown. As of August 30, Cursor has not announced credit handling, default-model changes, or migration support. Truell says the companies are “speaking to resolve this.”

Is this final, or could the shutoff be canceled? OpenAI calls November 12 a “proposed” shutoff and Cursor says talks are ongoing. Treat the date as real for planning purposes; be pleasantly surprised if it moves.

What’s the cheapest like-for-like replacement for GPT-5.6 Terra in Cursor? Claude Sonnet 5 at $2/$10 per million tokens — effectively the same price as Terra’s $2/$12 — or Composer 2.5 at $0.50/$2.50 through Cursor credits if you’re open to switching workflows to Cursor’s in-house model.

Sources

• Our decision on Cursor following its acquisition by SpaceX — OpenAI
• OpenAI announcement post — @OpenAI on X
• Michael Truell’s response (5% traffic figure) — @mntruell on X
• OpenAI to end model access to Cursor after acquisition by Elon Musk’s SpaceX — CNBC
• OpenAI will pull its models from Cursor due to SpaceXAI acquisition — Engadget
• OpenAI cuts off Cursor, citing Musk’s history of breaking contracts — The Decoder
• Anthropic vows to increase Claude compute for Cursor — Wccftech
• Personal API keys won’t cover every gap — XenoSpectrum
• GPT-5.6 Sol — Cursor Docs
• GPT-5.6 pricing: Sol, Terra, and Luna costs — CloudZero
• Claude model pricing — Anthropic Platform Docs
• Cursor pricing — official

Last verified August 30, 2026. This story is developing — the November 12 date is OpenAI’s proposed shutoff and talks between the companies are ongoing. Pricing and model availability change frequently; verify current state before purchasing.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe