On August 29, OpenAI published a blog post announcing it would wind down the contract that supplies its models to Cursor, with a proposed shutoff of November 12.
The stated reason was that SpaceX had just closed its $60B acquisition of Cursor’s parent company, and OpenAI didn’t trust Elon’s companies to honor its terms of service. Elon’s response was roughly that he couldn’t care less.
Nobody in that exchange was a developer using the tooling. But if you had GPT-5.6 wired into a Cursor code-review bot, a docs generator, or any CI hook that runs without a human watching it, you have about ten weeks to re-validate your workflow with a different provider.
The pitch of the entire API era has been that models are a utility - you rent them and you build on top.
What actually happened over the last twelve months is that a few hundred people in a few San Francisco offices, plus one directive out of DC, demonstrated that they can reach directly into a working developer’s stack and change what runs there. Not through malice, mostly - but through business decisions, contract disputes, roadmap slides, and one bad Friday afternoon at the Commerce Department.
The pattern, not the incident #
The Cursor/OpenAI situation is the loudest example, but it’s the sixth or seventh event in a year:
June 2025: Days after Bloomberg reported OpenAI’s $3B acquisition of Windsurf, Anthropic cut nearly all first-party capacity to Claude 3.x with under five days’ notice.
Jared Kaplan’s explanation was that it would be odd to sell Claude to OpenAI, which is honest and also a complete description of how vertical integration works. Windsurf ended up subsidizing Gemini 2.5 Pro at 0.75x to keep users from bailing.
August 16, 2026: Personal ChatGPT accounts quietly lost the ability to create new Custom GPTs. Business and Enterprise kept it, and existing GPTs still run with no sunset date announced, but if you were selling a product built on that format, you now can’t ship v2.
August 26, 2026: The OpenAI Assistants API shut down. /v1/assistants, /v1/threads, /v1/threads/runs, all gone, with no automated migration path.
Threads needed to be exported by hand and rebuilt as Conversations. Anything that wasn’t exported is effectively gone now. Zapier retired its ChatGPT (Assistants) steps the same day, which means a bunch of small businesses who never once thought of themselves as OpenAI API customers watched their automations die.
If you sort those by cause you get four separate choke points, each with a different owner. Model access is controlled by whoever signed the contract. Product surfaces are controlled by whoever owns the roadmap. Your state and history are controlled by whoever hosts it and decides whether export tooling is worth building. The law intervenes too, which we’ll get to next.
The one that made the news #
At 5:21pm ET on June 12, 2026, BIS (U.S. Bureau of Industry and Security) served Anthropic an export-control directive: suspend access to Claude Fable 5 and Mythos 5 for any foreign national anywhere, including Anthropic’s own foreign-national employees.
There is no way to verify citizenship in real time across hundreds of millions of accounts, so the only compliant move was to disable both models for every person on the planet. That’s the first time a commercial frontier model has been switched off by government order.
The trigger was a jailbreak demo from an Amazon researcher showing the model could find and exploit software vulnerabilities. Anthropic’s response was that the same capability is widely available from other models including GPT-5.5, and that if this standard were applied across the industry it would essentially halt all new model deployments from every frontier provider.
They weren’t wrong, but it didn’t matter. The models were down for 18 days, restored July 1 after Anthropic trained a classifier that blocks the technique in over 99% of cases, and had it validated by Commerce’s own standards body.
The takeaway here is that BIS possesses, and is willing to use, ad hoc emergency authority over model access, with no advance warning. So the precedent isn’t really about export controls. It’s that one alarming vulnerability report can unplug a deployed model in hours, globally, and your migration window is however long it takes someone to publish a blog post.
CNBC noted a criticism of the crackdown that is interesting: it handed valuable time to the Chinese open-source labs trying to catch up. Even the state’s biggest lever against a closed American lab ended up strengthening the open-weights world.
Why a file is different from a contract #
No open-weights model has ever been recalled this way, and the reason is boring. You can’t un-ship a checkpoint that’s already sitting on ten million hard drives, mirrored on Hugging Face, and seeded in torrents. Weights don’t have change-of-control clauses.
The receipts on quality caught up faster than most people expected. DeepSeek R1 landed in January 2025 with near-frontier reasoning, open weights, free, and Altman conceded OpenAI had been on the wrong side of history.
Seven months later OpenAI shipped gpt-oss under Apache 2.0, its first open-weight language models since GPT-2, with the 120b fitting on a single 80GB card and the 20b running on a 16GB laptop.
DeepSeek V4 Pro arrived in April 2026: MIT license, 1.6T-param MoE, around $0.43 in and $0.87 out per million tokens, roughly an order of magnitude under frontier API pricing, and you can host it yourself if you’d rather not trust anybody’s uptime.
The most compelling supporting detail in this whole argument for me, is that Cursor survived the OpenAI cutoff because it had already built its own model, Composer 2.5, on Moonshot’s Kimi K2.5 * open-weights* checkpoint.
Cursor’s own framing is an open base with proprietary intelligence, with 85% of compute going into additional training and RL on top. A company that just sold for $60B kept its flagship product alive on a foundation nobody can revoke. That’s not ideology, it’s risk management, and it worked on the first real test.
What this means for how you build #
Openness in this market is a competitive move as much as a principle. OpenAI shipped gpt-oss because DeepSeek forced its hand, and Cursor built on Kimi K2.5 because it needed a floor it couldn’t lose. That means the incentive to keep releasing weights is structural, not sentimental, and structural incentives outlast good intentions.
The asymmetry is in how these things fail. A contract ends on a change of control, a roadmap retires the surface your product sits on, and a directive can take a model offline worldwide before close of business, all of which happened in the last twelve months.
None of those mechanisms reach a checkpoint on your own disk, because there’s no legal instrument that gets to ten million copies of a file. So the practical version of the hedge is unglamorous: keep your prompts and evals somewhere you can repoint on an afternoon’s notice, and assume every hosted agent API is something you’ll be replacing eventually. That’s the lesson The Router took from the Assistants API shutdown, after it cost people workflows they’d spent years building. It’s also a huge part of the reason Kilo is open source and lets you switch models at any time: when a provider pulls out, that should be a dropdown.