# OpenAI cuts off Cursor: your model exit plan on Azure

OpenAI will stop supplying models to Cursor on 12 November 2026, invoking a change-of-control clause after SpaceX's $60 billion acquisition, and Cursor absorbed it because OpenAI carried only 5 percent of its traffic. For Azure buyers the lesson is a model exit plan: what Foundry's lifecycle policy and Microsoft's OpenAI licence actually guarantee, where Claude's EU Data Zone gap bites, the cost of a warm second source, and what DORA already requires.

- Published: 2026-09-02 · Category: Business & Strategy · Tags: OpenAI, Cursor, SpaceX, Microsoft Foundry, Vendor Risk, Model Portability, DORA, Claude, Azure OpenAI, Exit Strategy
- Author: Technspire AB, Stockholm (https://technspire.com)
- Canonical: https://technspire.com/en/blog/openai-cuts-off-cursor-your-model-exit-plan-on-azure

On 29 August 2026 OpenAI announced it had notified SpaceX that it will stop supplying models to Cursor, with a proposed shutoff on 12 November 2026. The trigger was a change-of-control clause. SpaceX closed its $60 billion all-stock purchase of Cursor's parent company Anysphere on 14 August, and OpenAI's custom agreement with Cursor gave it a limited window to cancel after an ownership change. Cursor's CEO Michael Truell says OpenAI models carry about 5 percent of the tool's traffic. That is the useful number in the story. One of the largest coding-assistant products in the world lost a frontier model line with roughly ten weeks' notice, over a dispute it had no part in, and shrugged it off because no single provider had ever been load-bearing. If you buy AI capability on Azure for a Swedish or EU organisation, the question worth asking this week is whether your own model supply could absorb a 12 November of its own.

## What happened, and when

| Date | Event |
| --- | --- |
| 16 June 2026 | SpaceX announces a definitive agreement to buy Anysphere, maker of Cursor, in an all-stock deal valued at $60 billion. |
| 14 August 2026 | The acquisition closes. Cursor is folded into the new SpaceXAI unit. |
| 29 August 2026 | OpenAI publishes its decision: it will wind down the contract supplying models to Cursor and will not provide future models, including Astra. |
| 12 November 2026 | Proposed shutoff date, described by OpenAI as the maximum notice its contract allows. |

OpenAI's statement is short and unusually direct. It says the company "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." It cites two episodes: Twitter breaking the terms of an OpenAI contract after Musk acquired it, and Musk's admission under oath earlier this year that xAI had violated OpenAI's terms of service. It then names the mechanism: "Our custom agreement with Cursor gives us a limited time window to cancel it after a change of control." OpenAI adds that its upcoming Astra model brings "a new level of accountability" and that it has decided "to hold the contract cancellation to the latest date we can while not providing future models to Cursor."

Cursor's response was measured. Truell wrote that OpenAI models serve about 5 percent of user traffic, that Cursor was one of OpenAI's first customers, and that the company had "trusted their platform to be neutral infrastructure." Anthropic's co-founder Tom Brown posted the opposite signal the same weekend: Cursor "has been a trusted partner of Anthropic since Sonnet 3.5" and Anthropic will "continue to increase compute to support Claude models in Cursor." Reporting by The Decoder adds that developers who want GPT models inside Cursor after November can still bring their own OpenAI API key.

## Read it as a supply-chain event, not a feud

Strip out the personalities and three mechanics remain. Each one applies to any enterprise that consumes frontier models through an intermediary, which in practice means every coding assistant, support bot and document tool you have licensed in the last two years.

**The change-of-control clause worked exactly as written, for the supplier.** Cursor's customers never saw that contract. Your SaaS vendors have equivalent agreements with their model providers, and those agreements almost certainly contain termination rights that are triggered by events on the vendor's side: an acquisition, a funding round with a competitor, a breach finding. None of those rights are visible from your end of the chain unless you ask.

**"Maximum notice" was ten weeks.** OpenAI framed 12 November as generous, and by the terms of its own contract it was. Ten weeks is also less than one sprint cycle for many enterprise change boards. Compare it with the 60 days Microsoft Foundry commits to before retiring a generally available model, or the "mandatory adequate transition period" that DORA requires financial entities to write into contracts for critical functions. Ten weeks is workable only if the second source already exists.

**Future models are withheld, which is a slow cutoff in disguise.** Even if the two companies settle on continued access to today's models, Cursor will not get Astra or whatever follows it. A frozen model line does not fail on a date. It fails when the rest of your stack moves on, when the provider retires the generation you are pinned to, or when a competitor ships on the newer model and your product falls behind. That failure mode is quieter than a shutoff and harder to put in a risk register.

Pareekh Jain of Pareekh Consulting put the lesson plainly to InfoWorld: "Enterprises can no longer assume that a model available through an AI coding platform today will always remain available. Acquisitions, contracts, competition or regulation can change this."

## What an Azure buyer actually holds

Buying models through Microsoft Foundry puts you in a different contractual position from Cursor's, but not a uniform one. It helps to be precise about what you hold, per model family.

**OpenAI models.** Under the October 2025 agreement between Microsoft and OpenAI, Microsoft's rights to OpenAI's models and products run through 2032. The April 2026 revision of that deal, as reported by The Decoder, made the licence non-exclusive and let OpenAI sell on other clouds, but left Microsoft's licence in place. That is a stronger footing than a reseller contract with a change-of-control window. It also covers only OpenAI.

**Models sold by Azure.** GPT models, DeepSeek V4, Cohere, Meta Llama, Mistral document and medium models, and Grok are sold and billed by Microsoft. They follow the Foundry lifecycle policy: a generally available model gets an 18-month clock set at launch, becomes unavailable to new customers at 12 months, and returns 410 Gone at 18. You get at least 60 days' active notice before retirement. Retirement dates are not extendable, and Microsoft reserves the right to an emergency retirement with shortened notice for compliance or security issues. Generally available models from Anthropic, DeepSeek, Fireworks and Mistral follow a 12-month lifecycle rather than 18.

**Models from partners.** Claude is the important case. Microsoft's documentation states that models from partners and community "are Non-Microsoft Products under the Product Terms." You buy Claude through Azure Marketplace, under licence terms and prices the provider sets, and Anthropic's own Supported Regions Policy may apply to availability. On the capability side the catalogue is rich: Claude Sonnet 5, Opus 5, Opus 4.8 and Haiku 4.5 are all deployable on Global Standard in Sweden Central. On the residency side there is a gap that matters for this post: Data Zone Standard for Claude exists only in the United States. In the EU tab of the partner-model region table, Data Zone Standard reads "Not available."

| Model family on Foundry | Counterparty | GA lifecycle | EU Data Zone | Sweden Central |
| --- | --- | --- | --- | --- |
| GPT-5.x (OpenAI) | Microsoft | 18 months | Yes | Yes |
| DeepSeek V4, Cohere, Llama, Grok | Microsoft | 18 months (DeepSeek 12) | Varies by model | Varies by model |
| Mistral (sold by Azure) | Microsoft | 12 months | Varies by model | Yes (Codestral, Ministral, Mistral Medium) |
| Claude (Anthropic) | Anthropic via Marketplace | 12 months | Not available | Yes, Global Standard |

The practical reading: a Foundry estate with GPT and Claude has two counterparties with different lifecycle clocks and different residency options. An exit plan that says "we can switch to Claude" is only true if your data-residency policy accepts Global Standard processing for that workload. If it demands EU Data Zone, your realistic second source today is another model sold by Azure, such as DeepSeek V4 or a Mistral model, and you should have evaluated it before you need it.

## Designing the exit plan

Cursor's 5 percent figure is the design target. The goal is a state in which losing any single provider costs you a routing change and a few days of quality regression, not a ten-week programme. Five components get you there.

### 1. Inventory what you depend on

Most organisations cannot list which model versions serve which workloads. The Foundry Models API answers that per subscription and region, and returns the lifecycle status and deprecation dates alongside the names. Pull it monthly and store the result.

```
az rest --method get --url \
  "https://management.azure.com/subscriptions/<sub>/providers/Microsoft.CognitiveServices/locations/swedencentral/models?api-version=2024-10-01" \
  --query "value[].{name:model.name, version:model.version, status:model.lifecycleStatus, retires:model.deprecation.inference}" \
  -o table
```

Join that list against your deployments, then against the applications that call each deployment. Add the SaaS layer by hand: for every licensed tool that embeds a model, record which provider it uses and whether you can bring your own key. That last column is the Cursor question in a single field.

### 2. Keep model names out of application code

Applications should call a deployment name that resolves in configuration, never a model identifier compiled into code. On Foundry the OpenAI-family models speak the Responses API and Claude speaks the Anthropic Messages API, so a genuine second source across providers needs an adapter that normalises the request and response shapes, including tool-call formats. The Microsoft Agent Framework and most agent SDKs already do this. If you wrote your own client, the adapter is the first thing to build, because everything else in the plan depends on being able to flip a switch.

### 3. Treat the eval suite as the exit plan's evidence

Without an evaluation suite you cannot say what a model swap costs you, and the change board will say no. Datadog published a useful public example on 26 August. It moved its internal coding workloads from Claude Opus to Claude Sonnet, cut AI spend by 36.7 percent, and measured an 8 percent loss of proficiency on its own workflows. It could state both numbers because it runs more than 140 evaluations a day against internal engineering tasks. That is the shape of evidence you want to have on file before a supplier announces anything: for each workload, the score on the primary model, the score on the standby, and the delta you are prepared to accept.

### 4. Keep the second source warm

A standby model you have never sent production traffic to is a plan on paper. Route a fixed slice of live traffic, 5 percent is a reasonable default, to the secondary at all times, and run the eval suite against both weekly. The cost is small. Take a workload of 400 million input and 40 million output tokens a month. At the list prices verified in our August cost posts, GPT-5.6 Terra on Azure is $2 per million input and $12 per million output, and Claude Sonnet 5 is $2 and $10.

| Line | Tokens per month | Cost per month |
| --- | --- | --- |
| Primary only (GPT-5.6 Terra, 100%) | 400M in / 40M out | $800 + $480 = $1,280 |
| Primary at 95% | 380M in / 38M out | $760 + $456 = $1,216 |
| Secondary at 5% (Sonnet 5) | 20M in / 2M out | $40 + $20 = $60 |
| Weekly evals, 500 cases, both models | ~6M in / ~2M out per week | ~$140 |
| **Total with warm standby** |  | **~$1,416** |

The warm standby adds around $136 a month, or about 11 percent, to a $1,280 workload, and almost all of that is the eval run rather than the traffic split. Scale the eval suite to your risk appetite. Set against the cost of an unplanned ten-week migration with engineers pulled off roadmap work, it is cheap insurance. Note the residency caveat from the previous section: if the primary runs on EU Data Zone, the Sonnet 5 slice runs on Global Standard, and your data protection officer needs to have approved that in advance for the specific data class involved.

### 5. Push the questions down your supply chain

For every SaaS tool that embeds a model, ask the vendor five things in writing. Which model providers does the product depend on? What notice does the vendor receive if a provider terminates, and what notice flows down to you? Can you bring your own key or your own Foundry deployment? Do your prompts, agent configurations and tool integrations survive a model change? Is there a change-of-control clause on either side, and who can invoke it? Cursor's customers would have had a very different August if they had known the answer to the last question in June.

## The Swedish and EU angle

**DORA already requires this from financial entities.** The Digital Operational Resilience Act, supervised in Sweden by Finansinspektionen, obliges banks, insurers and other financial entities to hold exit strategies for ICT services that support critical or important functions. Article 28(8) requires strategies that address "possible failure" of a provider and deterioration in service quality, with transition plans "enabling them to remove the contracted ICT services and the relevant data." Article 30(3) requires contracts for those functions to include "exit strategies, in particular the establishment of a mandatory adequate transition period." Article 28(7) gives entities termination rights when circumstances arise that could alter the performance of the functions provided, including material changes at the provider. A coding assistant is unlikely to be a critical function. A customer-facing support agent handling account queries may well be. Finansinspektionen has said ICT risk and DORA compliance are a supervisory priority for 2026, so the model layer of that support agent belongs in the register of ICT third-party arrangements with a named second source.

**Public procurement should ask for portability up front.** Swedish public-sector buyers running an upphandling for AI tooling can specify model portability as a requirement: a documented second source, bring-your-own-deployment support, and flow-down of provider termination notice. Those are cheap for vendors who have designed for it and impossible for vendors who have not, which is exactly what a requirements specification is for.

**The AI Act adds a wrinkle for high-risk systems.** From 2 December 2027, when the high-risk obligations begin to apply, a deployer that substantially modifies a high-risk AI system can take on provider obligations. Whether swapping the underlying model counts as a substantial modification will depend on the system and on guidance still to come, but it is a reason to run your eval suite and document its results before you switch, not after. The technical documentation you would need to show is the same documentation a good exit plan produces anyway.

## Checklist

- **1. Export the Foundry Models API for every region you use** and join it to deployments and applications. Add a column for each SaaS tool's model provider and bring-your-own-key status.
- **2. Classify each workload by data residency requirement.** EU Data Zone workloads need a second source sold by Azure; Global Standard workloads can use Claude in Sweden Central.
- **3. Name a standby model per workload** and record its eval score next to the primary's.
- **4. Move model identifiers into configuration** and put an adapter between your code and the provider APIs.
- **5. Route 5 percent of production traffic to the standby** and run the eval suite on both models weekly.
- **6. Send the five supply-chain questions to every vendor** that embeds a model, and file the answers with the contract.
- **7. If DORA applies, add the model layer to the ICT third-party register** with the transition period and termination triggers written into the contract.
- **8. Subscribe to Azure Service Health advisories for Azure OpenAI Service** so the 60-day retirement notices reach a person who owns the migration.

## Conclusion

OpenAI and SpaceX will argue about who broke faith with whom. Cursor's outcome is the part that transfers to your organisation: because no provider carried more than a small share of its traffic, a frontier vendor walking away became a routing change. Foundry gives Azure customers a firmer contractual base than Cursor had, with a licence through 2032 for OpenAI models and a published lifecycle with 60 days' notice for everything Microsoft sells. It does not remove the need for a second source, and for Claude it introduces a separate counterparty and a residency gap you have to plan around. Build the inventory, the adapter and the eval suite now, while the only deadline is one you set yourself.

## Sources

- [OpenAI: Our decision on Cursor following its acquisition by SpaceX](https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/)
- [CNBC: OpenAI to end model access to Cursor after acquisition by Elon Musk's SpaceX (29 August 2026)](https://www.cnbc.com/2026/08/29/openai-cursor-spacex-model-access.html)
- [CNBC: SpaceX to acquire the AI coding startup Cursor for $60 billion (16 June 2026)](https://www.cnbc.com/2026/06/16/spacex-spcx-cursor-acquisition-ipo.html)
- [Techzine: SpaceX completes acquisition of Cursor](https://www.techzine.eu/news/devops/143619/spacex-completes-acquisition-of-cursor/)
- [Michael Truell on X: Cursor's response to OpenAI's decision](https://x.com/mntruell/status/2093532254006063557)
- [Tom Brown on X: Anthropic will continue to increase compute for Claude in Cursor](https://x.com/NotTomBrown/status/2093541294027280657)
- [The Decoder: OpenAI cuts off Cursor after SpaceX acquisition](https://the-decoder.com/openai-cuts-off-cursor-after-spacex-acquisition-citing-musks-history-of-breaking-contracts/)
- [InfoWorld: Cursor customers will lose access to OpenAI coding models in November](https://www.infoworld.com/article/4216503/cursor-customers-will-lose-access-to-openai-coding-models-in-november.html)
- [Microsoft: The next chapter of the Microsoft-OpenAI partnership (28 October 2025)](https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/)
- [The Decoder: OpenAI and Microsoft rewrite their deal (April 2026)](https://the-decoder.com/openai-and-microsoft-rewrite-their-deal-no-more-exclusivity-no-more-agi-clause/)
- [Microsoft Learn: Foundry Models lifecycle and support policy](https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/model-retirements)
- [Microsoft Learn: Foundry Models from partners and community](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-from-partners)
- [Microsoft Learn: Foundry Models sold by Azure](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure)
- [Regulation (EU) 2022/2554 (DORA), Articles 28 and 30](https://eur-lex.europa.eu/eli/reg/2022/2554/oj)
- [Finansinspektionen: Om Dora](https://www.fi.se/sv/bank/it-risker-dora/om-dora/)
- [Datadog: How Datadog saves over $1 million each month by optimizing AI usage](https://www.datadoghq.com/blog/how-datadog-saves-money-by-optimizing-ai-usage/)

---

Technspire AB builds AI agents, Azure OpenAI solutions, and production web platforms for Swedish and EU enterprises. Book a call: https://calendly.com/technspire · hello@technspire.com · More articles: https://technspire.com/en/blog · Site overview for agents: https://technspire.com/llms.txt