OpenAI Cuts Cursor API on November 12: 5 Migration Routes That Actually Work in 2026
On August 29, 2026, OpenAI announced it will terminate the API contract powering Cursor's first-party integration, effective November 12, 2026. The trigger: SpaceX's $60 billion acquisition of Cursor parent Anysphere, which closed on August 14, 2026. OpenAI's stated reason, per Thibault Sottiaux: "It boils down to trust." The full OpenAI statement, reproduced verbatim by CNBC: "We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts."
The good news for the millions of Cursor users who never realized their editor was routing through OpenAI's contract: Cursor co-founder Michael Truell says OpenAI models account for about 5% of Cursor's user traffic. The other 95% is already Anthropic Claude (the dominant Cursor backend), Google Gemini, and a long tail of open-weights models routed through OpenRouter. The migration is therefore a story about completing the diversification Cursor's user base had already been doing, not about tearing down a single-vendor stack.
What this review covers: the exact November 12 cutoff mechanics (what breaks, what keeps working, what about future models), the 5 migration routes verified against live vendor pricing as of 2026-09-01, and the two industry precedents (Anthropic blocking Windsurf in June 2025, Anthropic revoking OpenAI's access in August 2025) that tell you how fast the dust settles.
Last verified: 2026-09-01. The cutoff date, the $60B SpaceX-Cursor deal size, the Truell 5% figure, the Tom Brown Anthropic X post, and the OpenAI pricing matrix are all primary-source-verified. If OpenAI extends or shortens the notice period, this article's recommendation set may need updating.
The Nov 12 cutoff: what breaks, what stays, what about new models
Three things change on November 12, 2026, and one very important thing doesn't change. From the OpenAI announcement, as reproduced by the-decoder.com and CNBC:
WHAT BREAKS on Nov 12, 2026
• Cursor's first-party OpenAI integration: GPT-5.6 family (Sol, Terra, Luna)
routed through Cursor's bundled API key stops working
• Cursor's access to future OpenAI model releases
(CNBC: "OpenAI also said it would not provide future models to Cursor
as it winds down the agreement")
WHAT STAYS WORKING after Nov 12
• BYOK (bring-your-own-key): users who already have an OpenAI platform
account and paste their personal API key into Cursor keep full access
to GPT-5.6 / future models — Cursor's BYOK code path is contractually
separate from the bundled integration
• Cursor's Anthropic integration (Claude Opus 5, Sonnet 5, Haiku 4.5) —
unaffected, Tom Brown publicly confirmed continued compute allocation
• Cursor's OpenRouter, Google Gemini, and open-weights routing — unaffected
NOTICE PERIOD
• 75 days from announcement (Aug 29 → Nov 12)
• Maximum notice period allowed under the contract's change-of-ownership clause
• Source: OpenAI blog (CF-walled) quoted second-hand via the-decoder.comThe single most consequential line is the third "WHAT STAYS WORKING" bullet. BYOK users are unaffected by the November 12 cutoff. If you already pay OpenAI directly for a GPT-5.6 API key and paste it into Cursor's settings, the editor keeps working exactly as it did yesterday. The contract being terminated is the one between OpenAI and Cursor's parent company, not the one between you and OpenAI.
The second most consequential line is the future-models lockout. Per CNBC: "OpenAI also said it would not provide future models to Cursor as it winds down the agreement." Translation: GPT-5.7, GPT-6, and any subsequent OpenAI release won't appear in Cursor's model picker for non-BYOK users. To get new OpenAI models after Nov 12, you must use BYOK (i.e., pay OpenAI directly) or route through a third-party like OpenRouter (which has its own OpenAI contract).
The 5 migration routes: verified pricing and trade-offs
I verified pricing against four primary sources on 2026-09-01: OpenAI's platform pricing page (platform.openai.com/docs/pricing), Anthropic's pricing docs (docs.anthropic.com/en/docs/about-claude/pricing), the OpenRouter models API (openrouter.ai/api/v1/models), and Cloudflare Workers AI pricing (developers.cloudflare.com/workers-ai/platform/pricing). Prices are USD per million tokens unless noted.
Route 1: BYOK with OpenAI GPT-5.6 Luna (cheapest OpenAI path)
If you specifically need GPT-5.6 quality and are willing to pay OpenAI directly, the cheapest sub-$1-per-million path is GPT-5.6 Luna on OpenRouter. From the OpenAI platform pricing matrix:
| Model | Input $/M | Cached Input $/M | Output $/M |
|---|---|---|---|
gpt-5.6-sol | $4.00 | $0.40 | $20.00 |
gpt-5.6-terra | $2.00 | $0.20 | $12.00 |
gpt-5.6-luna | $0.20 | $0.02 | $1.20 |
gpt-5.5 (<272K) | $5.00 | $0.50 | $30.00 |
Why pick Luna: it's the only GPT-5.6 tier under $1/M input. Use it for autocomplete, docstring generation, test scaffolding, and the high-volume low-stakes tasks where you'd otherwise burn GPT-5.6 Sol credits.
Why avoid Luna: it's the weakest GPT-5.6 variant — the input/cached ratio (10x cache discount) suggests aggressive quantization or distillation. For refactors, multi-file edits, and architectural reasoning, step up to Sol or Terra.
Setup cost: two minutes. Create an OpenAI platform account, add a payment method, generate a key at platform.openai.com/api-keys, paste it into Cursor's settings → Models → OpenAI API Key. Works from Nov 13 onward without changes.
Route 2: Switch to Anthropic Claude Opus 5 / Sonnet 5 (Cursor native)
Cursor already supports Anthropic Claude natively — and Anthropic explicitly signaled continuity. Tom Brown (Anthropic co-founder and Chief Compute Officer) on X, Friday night Aug 28/29: "Cursor has been a trusted partner of Anthropic since Sonnet 3.5. We'll continue to increase compute to support Claude models in Cursor and are excited for what comes next with them at SpaceX." Both CNBC and the-decoder.com reproduce the quote.
Verified pricing from Anthropic's pricing docs (2026-09-01):
| Model | Input $/M | Cached Input $/M | Output $/M |
|---|---|---|---|
| Claude Opus 5 / 4.8 | $5.00 | $6.25 | $25.00 |
| Opus 5 / 4.8 Fast mode | $10.00 | — | $50.00 |
| Claude Sonnet 5 | $3.00 | $3.75 | $15.00 |
| Claude Haiku 4.5 | $1.00 | $1.25 | $5.00 |
Why pick Opus 5: it's the strongest coding model Cursor currently supports natively, with 1M context window. Best for cross-file refactors, architectural decisions, and code review of complex PRs.
Why pick Sonnet 5: 60% cheaper than Opus on input, ~50% the cost on output. Sonnet has been Cursor's workhorse for two years; on coding benchmarks it sits within 8-15% of Opus while costing roughly half.
Setup cost: one minute. Add Anthropic as a provider in Cursor's settings, paste your Anthropic API key from console.anthropic.com. Cursor already routes Sonnet 4.5 today; Opus 5 / Sonnet 5 should appear in the model picker once Anthropic rolls out the model IDs.
Route 3: OpenRouter (universal fallback, 300+ models)
If you don't want to pick a vendor and just want one endpoint that handles every model, OpenRouter is the answer. It's been Cursor's third-party routing layer for two years and is unaffected by the OpenAI-Cursor contract dispute because OpenRouter has its own agreements with each underlying model provider.
Verified OpenRouter pricing on 2026-09-01 (top Cursor-friendly models):
| Model | Input $/M | Output $/M | Context |
|---|---|---|---|
openai/gpt-5.6-luna | $0.20 | $1.20 | 1,050,000 |
openai/gpt-5.6-luna:batch | $0.10 | $0.60 | 1,050,000 |
openai/gpt-5.6-sol | $2.00 | $10.00 | 1,050,000 |
anthropic/claude-opus-5 | $5.00 | $25.00 | 1,000,000 |
anthropic/claude-opus-5-fast | $10.00 | $50.00 | 1,000,000 |
google/gemini-3.7-flash | $0.75 | $3.75 | 1,048,576 |
google/gemini-3.5-flash-lite | $0.30 | $2.50 | 1,048,576 |
Why pick OpenRouter: one API key gives you access to GPT-5.6, Claude Opus 5, Gemini 3.7, and 300+ other models without per-vendor onboarding. Auto-failover when a model is down. OpenAI-compatible request format, so Cursor's existing code path works unchanged.
Setup cost: three minutes. Sign up at openrouter.ai, generate a key, paste into Cursor's OpenAI-compatible provider settings with base URL https://openrouter.ai/api/v1. Cursor 0.41+ supports the OpenRouter model picker directly.
Watch out: OpenRouter charges a small markup (typically 5-10%) on top of the underlying provider's list price. For high-volume workloads where you're paying 6+ figures a month, direct vendor contracts are cheaper. For individuals and small teams, the convenience premium is worth it.
Route 4: Cloudflare Workers AI (edge inference, 86 models)
If latency is your bottleneck — Cursor's Tab completion needs sub-200ms responses — Workers AI runs in 300+ cities and is competitive with frontier models on a $/performance basis. From the Workers AI pricing page (verified 2026-09-01): $0.011 per 1,000 Neurons on the standard tier, with 10,000 Neurons/day free on the free plan.
Cursor-relevant models on Workers AI (verified from the pricing page):
| Model | Input $/M | Output $/M | Context |
|---|---|---|---|
@cf/zai-org/glm-5.3 | $1.40 | $4.40 | 1,048,576 |
@cf/moonshotai/kimi-k2.7-code | $0.95 | $4.00 | — |
@cf/zai-org/glm-5.3-flash | $0.15 | $0.50 | 1,048,576 |
@cf/openai/gpt-oss-120b | $0.35 | $0.75 | — |
@cf/meta/llama-3.3-70b-instruct | $0.293 | $2.253 | — |
Why pick Workers AI: edge caching makes Tab autocomplete feel instant, 10K Neurons/day free for low-volume users, and the open-weights lineup (GLM-5.3, Kimi K2.7-Code, Llama 3.3 70B, GPT-OSS 120B) is competitive with closed-source peers at half the cost. Cursor already supports custom OpenAI-compatible endpoints, so pointing at https://api.cloudflare.com/client/v4/accounts/<id>/ai/v1 works.
Watch out: frontier models on Workers AI (GLM-5.3, GLM-5.2, Kimi K2.7-Code) require Workers Paid ($5/month) or prepaid AI Gateway credits — no free tier for these. Also, the rate-limit table for frontier models is still settling as of 2026-09-01.
Route 5: Google Gemini Flash (the sub-$1 frontier option)
If you want frontier-tier quality without frontier-tier pricing, Google's Gemini Flash family is the standout. All three tiers ship on OpenRouter and direct via Google AI Studio, all with 1M context, all under $1/M input:
| Model | Input $/M | Output $/M | Notes |
|---|---|---|---|
| Gemini 3.5 Flash Lite | $0.30 | $2.50 | Cheapest, batch: $0.15/$1.25 |
| Gemini 3.6 Flash | $0.75 | $3.75 | Batch: $0.375/$1.875 |
| Gemini 3.7 Flash | $0.75 | $3.75 | Newest, batch: $0.188/$0.938 |
Why pick Gemini Flash: Google has held the Flash family under $1/M input for three generations; the price-performance ratio is the best in the frontier class. Cursor supports Gemini via Google's API directly or via OpenRouter. The 1M context window handles even the largest monorepos.
Why avoid: coding-specific benchmarks put Gemini Flash behind Claude Sonnet 5 and GPT-5.6 Sol on agentic tasks. For "make this function work" prompts, Flash is fine. For "refactor this entire auth subsystem across 40 files," Sonnet 5 or Opus 5 will save you more time than Flash's price advantage costs.
The two precedents: how fast the dust settled
This isn't the first API contract to be terminated in the AI coding-tool space. Two precedents, both cited in the OpenAI Cursor announcement coverage, give a calibrated expectation:
Precedent 1: Anthropic blocks Windsurf (June 2025)
When reports surfaced that OpenAI was in talks to acquire Windsurf, Anthropic blocked Windsurf's access to Claude models. The result: Windsurf lost its default coding model overnight, pivoted to in-house models and third-party routing (OpenRouter + direct vendor agreements), and the user impact was contained because Windsurf users were already split across multiple backends. Anthropic's framing: "we cannot continue to provide access if the partner is being acquired by a competitor." Both the-decoder.com and CNBC confirm this is the precedent OpenAI is implicitly citing.
Settlement time: Windsurf had a working alternative within ~3 weeks. The bigger hit was Anthropic-Cognition's later acquisition of Windsurf's remaining team; the API cutoff itself was operationally absorbable.
Precedent 2: Anthropic revokes OpenAI's API access (August 2025)
OpenAI teams were using Claude for internal benchmarks ahead of the GPT-5 launch. Anthropic discovered this and revoked OpenAI's API access, framing it as a violation of the no-distillation clause. The action was symbolic — OpenAI had its own models — but it established that "trust" cuts both ways. Per the-decoder.com: "In August 2025, the tables turned on OpenAI when Anthropic revoked the company's API access because OpenAI teams had apparently used Claude for internal benchmarks ahead of the GPT-5 launch."
Settlement time: N/A — OpenAI didn't need to migrate, it just lost benchmark data. But the precedent for "API cutoffs as a competitive lever" was set.
Why the Cursor case is operationally different
The Cursor migration is bigger in user-count but smaller in technical-impact than Windsurf. Cursor has 75 days of notice (Aug 29 → Nov 12) versus Windsurf's reported ~24-hour cutoff. Truell's 5% figure means 95% of Cursor's stack is already multi-vendor. And unlike Windsurf (which lost its primary model), Cursor retains Claude (Tom Brown's X post is the explicit commitment) and its open-weights fallback (OpenRouter, Workers AI, direct Anthropic).
The realistic outcome: by Nov 13, 2026, Cursor will have either (a) negotiated a continued relationship with OpenAI for at least the GPT-5.6 Luna tier, or (b) defaulted to BYOK-only for OpenAI models while promoting Claude Opus 5 / Sonnet 5 and Gemini 3.7 Flash to first-class status in the model picker. Either way, the migration burden on individual users is small.
What I'd actually do today
Five concrete moves, ordered by effort vs. impact:
- Verify your OpenAI key is on a personal billing account, not Cursor's bundled one. Cursor → Settings → Models → OpenAI. If the field shows your own key starting with
sk-..., you're on BYOK and Nov 12 is a non-event. If it's blank or shows "managed by Cursor," you need to act. - Set up an Anthropic API key today. Two minutes at console.anthropic.com. Add to Cursor's Models settings. This is the most robust fallback — Tom Brown's X post means compute allocation is guaranteed.
- Sign up for OpenRouter. Free to register, $5 minimum to load credits. Even if you don't use it now, having the option to route to 300+ models with one key is the closest thing to "future-proof" against the next vendor dispute.
- Try Gemini 3.7 Flash for high-volume low-stakes tasks. Cursor → Settings → Models → Google → paste your Google AI Studio key. At $0.75/$3.75 per M tokens with 1M context, it's the most cost-effective frontier option for code completion and test scaffolding.
- If you self-host or have GPU budget, evaluate vLLM + GLM-5.3 / Kimi K2.7-Code. Both are Apache-2.0 / permissive licensed, both are competitive with closed-source peers on coding benchmarks, both run on 4× H100 80GB with INT8 quantization. Cursor's "OpenAI-compatible local" custom endpoint can point at your vLLM server. This is the only path that makes Cursor truly vendor-independent.
FAQ
Q: When does OpenAI's API cutoff for Cursor take effect?
November 12, 2026. Per both the-decoder.com and CNBC, this is the maximum notice period allowed under the existing contract's change-of-ownership clause. The 75-day window from the Aug 29 announcement is the longest of any comparable API cutoff in the AI coding-tool space (Windsurf got ~24 hours; the OpenAI benchmarks revocation was effectively immediate).
Q: Why is OpenAI cutting Cursor off?
SpaceX completed its $60 billion acquisition of Cursor parent Anysphere on August 14, 2026. Thibault Sottiaux (OpenAI) told CNBC: "We are making this choice because we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts." OpenAI separately cited the upcoming Astra model with "advanced cyber capabilities" as raising "a new level of accountability" for partner compliance — though CNBC did not independently confirm the Astra framing.
Q: How much of Cursor's traffic depends on OpenAI models?
About 5%, per Cursor co-founder Michael Truell's X post on Aug 29, 2026. Both CNBC and the-decoder.com quote Truell: "OpenAI models serve about 5% of Cursor user traffic, and we're speaking with the OpenAI team to resolve this." The other 95% is already Anthropic Claude (Cursor's dominant backend for two years), Google Gemini, and open-weights models routed through OpenRouter.
Q: Can existing Cursor users keep using GPT-5.6 after Nov 12?
Yes, via BYOK. Per the-decoder.com: "Users who access GPT models through Cursor will still be able to use their own OpenAI API keys, and OpenAI will continue to provide access through its IDE extensions for Cursor." The contract being terminated is between OpenAI and Cursor's parent company, not between you and OpenAI. The catch: per CNBC, "OpenAI also said it would not provide future models to Cursor as it winds down the agreement." New OpenAI model releases won't appear in Cursor's first-party model picker for non-BYOK users.
Q: What is Anthropic's response?
Tom Brown, Anthropic co-founder and Chief Compute Officer, posted on X Friday night (Aug 28/29, 2026): "Cursor has been a trusted partner of Anthropic since Sonnet 3.5. We'll continue to increase compute to support Claude models in Cursor and are excited for what comes next with them at SpaceX." Both CNBC and the-decoder.com reproduce the statement verbatim. The phrasing is a direct counter-positioning to OpenAI's "trust" framing — Anthropic is signaling that it sees the SpaceX-Cursor deal as continuity, not rupture.
Q: Which Cursor migration routes are most cost-effective today?
For most users, the answer is Anthropic Claude Sonnet 5 ($3/$15 per M) as the primary replacement and OpenRouter as the universal fallback. For high-volume Tab completion, GPT-5.6 Luna at $0.20/$1.20 per M via BYOK is the cheapest OpenAI-compatible path. For bulk test scaffolding and doc generation, Gemini 3.5/3.6/3.7 Flash at $0.30-$0.75 per M input is the most cost-effective frontier option. For self-hosters with GPU budget, GLM-5.3 / Kimi K2.7-Code on Workers AI or local vLLM is the only vendor-independent path.
Q: How does this compare to past API cutoffs (Windsurf, GPT-5 benchmarks)?
Two precedents frame this. June 2025: Anthropic blocked Windsurf from accessing Claude after reports OpenAI wanted to acquire Windsurf. Settlement time ~3 weeks. August 2025: Anthropic revoked OpenAI's API access after OpenAI teams used Claude for internal benchmarks. No migration needed (OpenAI had its own models) but the precedent for "API cutoff as competitive lever" was set. The Cursor case is the inverse — the model vendor is cutting the consumer off, not a competitor doing the cutting — and the 75-day notice period is the longest of the three precedents.
Bottom line
OpenAI's API cutoff for Cursor is a symbolic escalation in the ongoing Musk-OpenAI feud, not a technical catastrophe for Cursor users. The 5% traffic dependency, the 75-day notice period, the BYOK escape hatch, the Tom Brown / Anthropic continuity commitment, and the mature OpenRouter / Workers AI / Gemini fallback routes all add up to a manageable migration.
The bigger structural signal is that API contracts between model vendors and consumer apps are now a competitive lever. The Windsurf precedent (June 2025), the OpenAI benchmarks revocation (August 2025), and now the Cursor cutoff (November 2026) form a clear pattern: any AI coding tool that depends on a single vendor's goodwill is one acquisition announcement away from losing access. Multi-vendor routing (OpenRouter) or self-hosted open weights (GLM-5.3, Kimi K2.7-Code) are no longer "nice to have" — they're the only durable architecture.
If you're starting a new AI coding project today, the lesson is clear: design for model portability from day one. Use OpenRouter or Workers AI as your routing layer, keep two vendor contracts live at all times, and never let a single model provider become a single point of failure for your dev workflow.
Last verified: 2026-09-01. Primary sources: the-decoder.com (Matthias Bastian, Aug 29 2026), CNBC (Lora Kolodny, Aug 29 2026), OpenAI platform pricing, Anthropic pricing docs, OpenRouter models API. The OpenAI blog URL openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex was Cloudflare-walled at fetch time; OpenAI statements are quoted second-hand via the-decoder.com and CNBC, both of which independently reproduce the same text. Reuters, Forbes, and Techzine URLs also returned CF-walled responses in this environment but the $60B deal size is corroborated by CNBC's direct reporting.