Nobody's model supplier is neutral infrastructure
SpaceX closed its $60 billion purchase of Anysphere on 14 August. OpenAI gave notice fourteen days later, using a clause about who owns the company rather than anything about how it behaved.
Status as of 3 September 2026: the 12 November shutoff is proposed, not settled. Cursor says talks with OpenAI are ongoing. This post gets refreshed on 12 November, or sooner if either side announces a change.
By Finley Jones, co-founder and CCMO, Taskpool International Ltd. Published 3 September 2026.
On 28 August, OpenAI told SpaceX it was winding down the contract that supplies its models to Cursor, with a proposed shutoff date of 12 November 2026. Nothing had broken. No engineer shipped a bug, nobody blew through a rate limit, no benchmark moved. The cap table changed and a contractual right activated.
The easy read is a Musk-versus-Altman story, and Musk supplied the material for it. He posted that he "couldn't care less" and then insulted OpenAI's leadership by name.
But a model supply relationship running for nearly four years ended over a change-of-ownership provision, and that deserves more attention than the insults. If you are building a product on weights you do not own, the clause is the part that transfers to your situation. The feud does not.
What happened, in order
SpaceX secured an option on 21 April 2026 giving it the right to acquire Anysphere for $60 billion, or to walk away for roughly $10 billion covering a break fee and joint work. It exercised the option on 16 June in an all-stock deal, four days after pricing the largest IPO on record: 555.6 million shares at $135, raising $75 billion at a $1.77 trillion valuation. The stock closed its first session up 19%, putting the company above $2 trillion. The Anysphere transaction closed on 14 August, converting Anysphere stock into roughly 389 million SpaceX Class A shares and folding Cursor into the SpaceXAI division.
Anysphere's annualised revenue reached about $4 billion by early June, roughly $2.6 billion of it enterprise, having passed $2 billion in February. Four MIT classmates started the company in 2022, and Forbes put each cofounder near $2.7 billion at close. Before SpaceX, Cursor had rebuffed two approaches from OpenAI and Microsoft had looked at an acquisition without submitting a bid.
Two things in OpenAI's statement matter more than its tone. It was explicit that this was a change-of-control decision and made no accusation that Cursor had broken any rule. And the model being protected was Astra, OpenAI's upcoming release, which will not reach Cursor's model picker at all: the company said it would hold cancellation to the latest date the contract allows while supplying no new models in the interim. Its stated reason was that it could not be confident SpaceX would operate within its terms of service, pointing to the Twitter contract and to Musk's admission under oath in April that xAI had distilled OpenAI outputs.
The 5% number is doing a lot of work
Cursor CEO Michael Truell replied with a figure: OpenAI models serve about 5% of Cursor user traffic. He added that the companies were talking.
A traffic share published by an interested party mid-dispute deserves scepticism. Truell has not published a provider-by-provider breakdown, and "user traffic" is not "revenue" or "enterprise seats" or "sessions that mattered". Treat it as the number Cursor chose to release, not as an audited exposure figure.
The structural claim underneath it survives the scepticism, though, because the receipts are public. Cursor spent the previous year demoting frontier models from dependency to line item.
Composer 2.5 shipped on 18 May 2026, two months after Composer 2. It is built on Moonshot AI's Kimi K2.5, an open-weight mixture-of-experts checkpoint with roughly a trillion total parameters and about 32 billion active per inference. Cursor says around 85% of the model's total compute went into its own continued pretraining and reinforcement learning, on 25 times more synthetic coding tasks than Composer 2.
The results were close enough to matter. On SWE-Bench Multilingual, Composer 2.5 scored 79.8% against Claude Opus 4.7's 80.5%. On Terminal-Bench 2.0, 69.3% against Opus 4.7's 69.4%. It is not parity everywhere. GPT-5.5 holds 82.7% on that same Terminal-Bench figure, thirteen points clear, and Cursor conceded that some of what it improved, effort calibration and communication style, is not captured by public benchmarks at all. Independent replication on a unified scaffold has not happened. Price is where the gap opens the other way: standard Composer 2.5 runs at $0.50 and $2.50 per million input and output tokens against Opus 4.7's $5 and $25, and Cursor's own effort curve puts it near 63% on CursorBench at under a dollar of average cost per task. Both of those are Cursor's numbers about Cursor's model.
So OpenAI's withdrawal took about 5% of Cursor's traffic and none of its core capability. The fallback was already built, already cheaper, already the default for the workload it was designed for.
The base model turned out to be the cheap part to copy. The floor under a competent coding agent in 2026 is an open-weight checkpoint plus a post-training budget. Cursor did not need to train a frontier model from scratch to make its suppliers substitutable. It needed a public base and a serious RL pipeline. The economics of that trade get a fuller treatment in You Cannot LoRA Your Way to Composer.
Nobody diligences the cap table
Engineering teams evaluate model providers on latency, context window, rate limits, benchmark scores, and price per million tokens. Those are the fields in every comparison table ever written on this subject.
Almost nobody evaluates the acquisition scenarios of their provider's counterparties, or reads the termination rights in the contract that supplies them. Model supply is a legal artefact as much as a technical one, and here the risk sat in the cap table rather than the API.
OpenAI's own explanation makes the mechanism explicit: it relies on custom contracts with large partners to enforce terms of service and maintain safety at scale. Custom contracts contain change-of-control provisions. Cursor's contained one that opened a limited window to cancel when ownership changed, and OpenAI used it.
The corollary for smaller teams is worse, and this is the part I would want a founder to take away. Most are on self-serve terms, which carry no negotiated notice period at all. Cursor got until 12 November because it was big enough to have a bespoke agreement. A startup in the same position gets whatever the standard terms say, which is usually far less, and finds out by reading a status page.
Supplier, competitor, landlord
The neat mental model of AI infrastructure has compute at the bottom, model labs above it, applications on top. This deal shows how thoroughly that has collapsed.
Anthropic has been buying compute from SpaceX since May, including capacity on Colossus, according to The State of AI. SpaceX now also owns one of the largest distribution surfaces for Anthropic's models. Within a day of OpenAI's announcement, Anthropic cofounder Tom Brown posted on X that Cursor had been "a trusted partner of Anthropic since Sonnet 3.5" and that it would increase compute behind Claude in Cursor. Replit's CEO replied by reminding him about Windsurf.
Which is the fact I have been holding back. Anthropic cut Windsurf's Claude access in June 2025 on less than five days' notice while OpenAI was circling that company, and in January 2026 it cut off xAI's engineers, having found them because they were coming in through Cursor. Anthropic's terms are broader than OpenAI's on this point, barring customers from building competing products at all. So the company now describing Cursor as a trusted partner has the widest contractual grounds of any frontier lab to do what OpenAI just did, and has used them twice, against Cursor's new owner among others.
Both postures are defensible on their own terms, which is the difficulty. Neither is neutrality. OpenAI enforced against a supplier relationship it no longer trusted and lost the distribution. Anthropic kept the distribution and now enforces terms of service against its own landlord, or does not.
There is a second thing the compute story explains better than editor market share does. Cursor said in May, before the option was exercised, that it was training a significantly larger model from scratch with SpaceXAI, using ten times the total compute of Composer 2.5 against Colossus 2's million H100-equivalents. It had already drawn on tens of thousands of xAI chips for Composer 2.5, and two senior Cursor engineers had already left for xAI. Read forward from that, the acquisition ratified a dependency that existed before anyone signed anything. Cursor made its model suppliers substitutable by making itself dependent on a compute supplier, and then the compute supplier bought it. Whether that is a better position is not obvious to me.
How to build against provider change-of-control risk
"Avoid vendor lock-in" has never been useful advice, because nobody turned it into a task. These are tasks.
Put model selection behind an interface from day one. A better model will appear, but that is the smaller reason. The bigger one is that a supplier can be removed by an event with nothing to do with product quality. Provider-specific behaviour belongs in adapters, not in prompt strings scattered across a codebase. The cost is real: an abstraction layer means you ship later than a team that hardcodes one provider, and you will lose some provider-specific features to the lowest common denominator.
Maintain an eval suite you can re-run in an afternoon. Cursor could quantify its exposure and compare Composer against Opus on named benchmarks because it had built the measurement apparatus first. Most teams substituting one model for another are running on vibes and a few manual spot checks, which means they cannot tell a supply problem from a quality regression.
Know the delta on the fallback before you need it. A second provider configured but never measured is not redundancy. The useful artefact is a number: what changes in quality and cost, on your workload, on your evals, when traffic moves.
Read the termination and change-of-control terms in whatever agreement you are actually on. For most teams that is the self-serve terms, and it takes twenty minutes. You will probably find no notice period. Knowing that is still worth the twenty minutes, because it tells you how much fallback you need rather than how much you would like.
Treat any capability that exists in exactly one place as product risk. If a feature only works because of one provider's model, it has the same availability guarantee as the commercial relationship behind it. Cursor's answer was Composer on an open base. The cheaper version is to pin to open-weight checkpoints available through several hosts, which costs you the frontier and buys you the option.
The strongest case against all of this: abstraction layers are a tax paid every sprint against a risk that fires once every few years, if ever, and the teams that hardcoded GPT-4 in 2023 and shipped are mostly still fine. If your product would survive a three-month migration, that case is probably right.
What the question is now
The competitive question in AI products has changed shape. Which model is best is only part of it. The rest is who owns the supplier, what the contract permits them to do, and how quickly the product absorbs their departure.
Cursor took a supply cut from OpenAI and the story was over in a news cycle. The relationship mattered, and Cursor had spent a year and a large compute budget making sure it did not have to. None of that work was visible until 28 August, when it turned out to be the reason the company could shrug. What is not visible yet is whether the 12 November date holds at all, given that the two sides are still talking, and I do not have a read on it.
The 5% figure is the number in this post I am least confident in. If you have a better estimate of Cursor's provider mix, or you have actually read a frontier lab's change-of-control language and can say what the standard notice period is, I want it: @finjonesceo.