When Your AI Provider Says No

Introduction

You rolled out an AI tool to your team a year ago. Maybe it was ChatGPT Team for the office, Copilot for the two people who write code, or an AI feature that showed up inside the CRM you already pay for. Whatever it was, it stuck. People use it every day, and if it disappeared tomorrow, someone would notice before lunch.

On August 29, 2026, OpenAI told Cursor opens a new window , one of the most widely used AI coding tools, that Cursor’s access to OpenAI models ends on November 12. There was no outage, no price change, and no wrongdoing on Cursor’s side. Cursor had just been bought by SpaceX, and OpenAI decided it could not trust the new owner to honor its terms. Cursor’s CEO wrote that they had “trusted their platform to be neutral infrastructure for our business.”

That sentence is why this post exists. Cursor is a good example, and this post is not really about Cursor. It is about what it means to run a business on tools whose provider can decide, for reasons that have nothing to do with you, that you can’t use them anymore. We ran the exercise on ourselves at OmbuLabs: where AI actually sits in the business, which company is underneath each tool, what stops if that company says no, and which of those gaps are worth paying to close. What follows is that exercise.

What happened to Cursor, and what didn’t

The timeline is short. SpaceX agreed to buy Cursor in June for about $60 billion in stock, and the deal closed on August 14. Fifteen days later, OpenAI announced that it would stop serving its models to Cursor on November 12, giving Cursor roughly ten weeks and the option to pick an earlier date. OpenAI’s stated reason was earlier disputes with Elon Musk’s companies, including the admission that xAI had trained on OpenAI outputs.

The way Cursor used those models did not change. Its owner did, and OpenAI made a decision about that owner. OpenAI’s help center opens a new window lists three ways to keep using its models in Cursor after that, your own API key, OpenAI’s Codex extension, or a gateway such as Amazon Bedrock or Azure, and all three mean the thing you were paying Cursor for is now something you set up and pay for yourself.

Cursor’s CEO, Michael Truell, said opens a new window OpenAI models accounted for about 5% of Cursor’s traffic. Most of its users were already on Anthropic’s Claude or Google’s models, both of which keep working, and Anthropic said publicly opens a new window it would add capacity for Claude in Cursor. Cursor will be fine.

It is worth being honest about why. Cursor did not spread its usage across model providers as a continuity strategy. It offered several models because its users wanted the choice, and those users drifted toward Claude because they preferred it for coding. The resilience was real, but nobody planned it, and that is the part you can’t count on for your own business.

The other ways a provider says no

Cursor is not the first coding tool this has happened to, and OpenAI is not the only provider that does it. In June 2025, while OpenAI was reportedly buying Windsurf, Anthropic cut off opens a new window nearly all of Windsurf’s direct access to Claude models with, in Windsurf’s words, less than five days of notice. Windsurf said it had wanted to pay for the full capacity. It is the same shape as Cursor, with a different provider doing the cutting and about a tenth of the notice.

Acquisitions are one trigger, and geography is another. In June 2024, OpenAI emailed opens a new window developers that it would block API traffic from countries it doesn’t support starting July 9, about two weeks out. In September 2025, Anthropic updated its terms opens a new window to refuse service to any company more than 50% owned by entities headquartered in places like China, wherever that company actually operates. A business in Singapore or Germany with the wrong parent company lost access without moving anywhere.

Retirements are the quiet version. On January 29, 2026, OpenAI announced opens a new window it would remove GPT-4o and three other models from ChatGPT on February 13, noting that only 0.1% of users still picked GPT-4o on a given day. Business and Enterprise plans kept it inside custom GPTs until April 3. If your team built a workflow around one model’s particular behavior, a retirement breaks it as thoroughly as a ban does, just more politely.

Across all of these, the notice ranged from ten weeks to five days, none of the affected customers had done anything wrong, and both of the providers most businesses use today, including the one we use most, have done it at least once.

The inventory

The exercise itself is not complicated, and a spreadsheet is enough. Four columns: the tool, the company whose model runs underneath it, who in the business uses it, and what stops if it goes away. Here is what a first pass might look like for a small accounting firm, made up for the example:

The first column is easy, and the third and fourth are mostly a matter of asking around. The second column is the one people skip, because for most tools the provider underneath is nowhere on the invoice. Each of those tools is somebody else’s product, calling a model provider you never chose, and it is not unusual for several of them to be calling the same one, which means a single “no” can take out three rows at once.

Finding out is usually possible. Most software vendors publish a list of subprocessors, the other companies they send your data to, because privacy regulations require it, and the model provider tends to appear there. Failing that, the vendor’s documentation or a direct question to your account manager will do. It is the same question we put on the list for safety reasons in The Guardrail Question to Ask Any AI Vendor opens a new window , asked this time for continuity reasons.

The one answer to be wary of is “we use multiple providers.” It is a good sign, but ask which provider handles the feature you actually depend on. Cursor also used multiple providers, and its OpenAI users still lost their model.

What breaks, and what a fallback costs

With the fourth column filled in, most rows sort themselves into one of three piles. Some are an annoyance: people lose a shortcut and go back to doing the thing the way they did it eighteen months ago. Some slow a team down in a way that shows up in the numbers, a support queue that grows or a sales team that stops logging calls. A few stop something that makes money or keeps you compliant, like a chatbot that is the only after-hours channel or a document pipeline a client deadline depends on. Be strict about which pile each row belongs in, because the temptation is to put everything in the third one, and that is how you end up paying for fallbacks nobody will use.

For the first pile, do nothing. Diversifying has a cost, and for a tool people would grumble about losing and then get on with their day, the cost is not worth it. For the second pile, the cheapest insurance is usually not a second provider. It is making sure the things that live inside the tool also live somewhere else: the prompts and instructions people have refined over months, the kind of shared knowledge Fiona described in AI for Small Business Starts With People opens a new window , the document templates, the custom assistants someone built, exported into a folder you control. A workaround like bringing your own API key gives you the model back, but whatever was built inside the tool around it, saved conversations, custom assistants, months of refined instructions, does not move with you.

For the third pile, a real second option is justified, and it is usually cheaper than it sounds: a second seat with another provider for the two or three people who depend on the tool most, tested once a quarter so someone knows it works, and a written version of the manual process, kept current. What is expensive is running two providers for everything at once, because the outputs differ and someone has to check both. We would not recommend that for a business without an engineering team, and even for us it only makes sense on a few rows.

Our own inventory

We are a software company, so our list looks different from the accounting firm’s, but the columns are the same.

The biggest row is Claude Code, from Anthropic. Most of our engineers use it daily, and our blog writing process, the one we described in AI Assistant for Our Blog Writing Process opens a new window , is a set of Claude Code skills on top of a server we built. If Anthropic said no tomorrow, engineering would slow down noticeably for a few weeks and the blog workflow would have to be rebuilt for a different agent. That puts it in the second pile, and we have decided to accept the dependency. The parts we own, the server and the skill files, are plain code and plain text that other agents can read, so what we would actually lose is the agent itself and the habits built around it, and no second provider gives us those back. What we do instead is keep the skills and the server portable and avoid building on features only Claude Code has.

The same server uses Voyage AI for embeddings, the numbers that make the search over our posts work, as we covered in A Blog MCP Server with FastMCP and pgvector opens a new window . Voyage was bought by MongoDB opens a new window in February 2025, before we chose it, so that row already carries an ownership change. If it went away, re-embedding a few hundred posts with another provider is a script and an afternoon. That is the first pile.

The newsletter assistant we wrote about in 2025 ran on OpenAI models. It has since become a marketing assistant, folded into the same platform as the blog server, and it runs on Anthropic models now. The row moved from one provider to the other without anyone treating it as a provider decision, and our list is more concentrated on Anthropic than it was a year ago.

Two smaller rows point elsewhere. Our Rails Superhero Card Generator opens a new window , the small app that turns a photo and a few skills into a superhero card, uses Google’s Gemini image models. A report generation tool we use now and then runs on OpenAI models. Both would be missed for about as long as it takes to notice, and neither gets a fallback.

The rows we expected to be the hardest to fill, the AI features inside the software we pay for, turned out to be mostly empty. GitHub, Slack, and Jira all ship AI features, and we barely use them. Our AI usage is our own tooling and Anthropic’s Claude products, which makes our inventory shorter than most and more concentrated than most. Client work stays off this list, for their sake rather than ours.

Conclusion

Cursor lost a provider it had trusted to be neutral, and it came out fine because its users had already spread themselves across three providers, for reasons that had nothing to do with resilience. The other cases we looked at came with notice periods from ten weeks to five days, both of the big providers were on the giving end at least once, and none of the customers involved had done anything to deserve it.

The exercise that follows from that is a four-column inventory, with the provider column being the one worth the effort, a sort into three piles, and a fallback only for the rows in the third pile. For our own list, the result was one accepted dependency, several rows that need nothing, and a concentration on a single provider that is higher than it was a year ago and that we are choosing to live with for now.

The caveat is that an inventory is a snapshot. Ours changed in under a year without anyone deciding to change it, as one tool became another and switched providers on the way. Providers change their terms faster than anyone updates a spreadsheet, so the useful habit is a short review whenever something new gets rolled out, or whenever a headline like Cursor’s shows up.

Not sure which provider sits under the tools your team uses every day, or which ones would actually hurt to lose? We can help you find out, so let’s talk opens a new window .