The AI Tool You Use Does Not Own the Model It Runs On

If you want to know which model your AI tool actually runs, the answer is rarely on the pricing page — it lives in four quieter places: the vendor’s model page, its product changelog — the dated list of what changed and when — its subprocessor list, and the terms or data-processing agreement behind your account. Late on Friday, August 28, 2026, OpenAI announced it would wind down the contract that supplies its models to the AI coding tool Cursor. Nobody who pays for Cursor signed that contract. That is the arrangement this post explains.

Here is the finding, up front: most AI products you pay for are resellers. The company whose logo is on your invoice licenses a model — the underlying AI system that actually generates the answers — from a lab like OpenAI, Anthropic, or Google, under a commercial agreement between those two companies. That agreement can change or end for reasons that have nothing to do with your account, your satisfaction, or your usage. When it does, your data almost never disappears — but what answers your prompts on Monday morning might be different from what answered them on Friday.

This is not a directory of which tool uses which model — any such table rots within a month, and the vendors’ own pages are always fresher than a blog post. Instead, this is the durable version: where a non-engineer looks, what each page can and cannot tell you, what a routine model swap actually changes (very little), what a real shutdown looks like (GitHub gave the textbook example in July 2026), and a ten-minute check you can run against any AI tool your business pays for.

Start with the distinction that makes everything else make sense. The tool is the product you log into — the editor, the workspace, the assistant in your browser. The model is the AI system that actually reads your prompt and writes the answer. For many of the tools businesses pay for, those are made by different companies, connected by a licensing contract.

You do not have to take that on faith — vendors say it themselves, on their own pages. Notion’s help documentation names Anthropic and OpenAI as suppliers of the models behind Notion AI. DuckDuckGo’s help page for Duck.ai lists every model at every tier by vendor — Anthropic, Mistral, and OpenAI among them, as published in August 2026. Raycast’s own comparison page names models from eleven different vendors. And Cursor’s models documentation states plainly that it supports frontier models from OpenAI, Anthropic, and Google — both pages as published in August 2026.

Two things follow from the rental arrangement, and they pull in opposite directions. First, the reassuring one: renting is normal, often good for you, and usually invisible — it is how a note-taking app gets a frontier-quality model without spending billions training one. Second, the uncomfortable one: the supply contract is between two companies you never see, negotiated on terms you never read, and it can end. When DuckDuckGo swaps one listed model for a newer one, its help page simply changes — notably, that page carries no visible last-updated date and no statement about how users are told when a swap happens. That silence is itself part of the pattern this post is about.

How common is this arrangement? Honestly: we found no reliable count, and we will not invent one. What can be said is qualitative — across productivity software, coding tools, browser assistants, and search, the vendors that disclose their supply chains routinely name outside labs as the source of their models. If you pay for an AI feature and have never checked who supplies it, the safe working assumption is that someone does.

Most vendor-risk writing lumps every AI supply change into one scary bucket. That is wrong, and the error matters, because the correct response to each event is different. There are two distinct things that can happen to the model behind your tool, and only one of them should ever put you on a clock.

The swap side of this table deserves more reassurance than it usually gets. Across every vendor changelog reviewed for this post — Notion’s, GitHub’s, Raycast’s — a model swap or addition is described in terms of new capability or continuity. None of the sources we examined describes a user’s prompts, outputs, or history being deleted or made inaccessible as a result of a model change. Swaps even happen live, mid-session, as a reliability feature: in September 2024, Cursor users reported seeing an in-product message reading “Anthropic is experiencing a partial outage right now. To avoid downtime, we’ve temporarily fallen back to an OpenAI model” — a substitution whose stated purpose, in Cursor’s own wording, was to keep the product working.

Here is the same distinction as a table you can hold onto — what actually changes in each event class, with the documented evidence behind each cell.

Read the columns, not the rows: everything in the swap column says “you are fine.” Everything in the retirement column says “you have a deadline.” The rest of this post is about telling the two apart early — which means knowing where each kind of change gets announced.

The four kinds of page below differ in how current they are, how precise they are, and how binding they are — which is why you check more than one. A quick vocabulary note before the table: a changelog is the vendor’s dated, running list of product changes (sometimes called release notes), and a subprocessor list is the formal roster of outside companies a vendor allows to touch customer data — more on that one in the next section.

The changelog deserves a special word, because it is where both event classes from the previous section show up — with dates. When GitHub announced on October 29, 2024 that Copilot was adding Anthropic’s Claude 3.5 Sonnet, Google’s Gemini 1.5 Pro, and OpenAI’s o1 models, that dated post marked the moment a single-supplier product became a multi-supplier one. Then its October 23, 2025 changelog entry retired a wave of models from all three suppliers, mapped each to a successor, and told users both “Please update your workflows and integrations to use supported models” and “No action is required to remove the deprecated models.” That pair of sentences is the swap-versus-shutdown distinction in the vendor’s own voice: builders had homework, everyday users had none. If your team runs automations against model APIs directly, our guide to surviving model deprecation calendars is the engineering-side companion to this section.

Of the four surfaces, one stands above the rest, and it is the one almost nobody outside a legal or procurement team ever reads. A subprocessor is any outside company a software vendor allows to handle its customers’ data — hosting providers, email services, and, crucially for this post, whoever hosts the AI model your text is sent to. Under Europe’s GDPR (Article 28), a vendor must have its customers’ authorization before engaging a subprocessor — either specific consent, or a published list customers can review and object to as it changes. That legal mechanism is why nearly every serious SaaS vendor maintains a public subprocessor page, usually on a “trust center” subdomain.

This is what makes the subprocessor list the most reliable public place to see whose model a tool sends your text to: it is a contractual disclosure, not marketing copy. A marketing page can say “powered by leading AI”; a subprocessor list has to name names, or at least categories a customer can object to, because customer contracts hang off it. Marketing pages answer to the brand team. Subprocessor lists answer to the contract.

And when the two disagree, you learn something. Notion — to its credit, one of the more transparent vendors examined here — describes its model suppliers at two different levels of precision on two of its own pages: the help article quoted in section 01, and its formal subprocessor table, as published in August 2026:

We are not smoothing that discrepancy over, because it is the point: the same vendor, describing the same supply chain, is precise where a contract demands it and general where prose allows it. That is exactly why the checklist in section 08 sends you to the trust center first and the help docs second. One caution in the other direction — Notion’s subprocessor page, as viewed for this research, carries no visible change-date stamp, so treat any roster you read there as “as of the day you looked,” and look again when a contract renewal is on the table.

You now know the rare, serious event is the access-layer retirement. So what does a well-run one look like? July 2026 supplied the template. GitHub Models was GitHub’s surface for trying and calling AI models directly — a playground, a model catalog, an inference API, and a bring-your-own-key option. On July 1, 2026, GitHub’s changelog announced, in plain words: “GitHub Models will be fully retired on July 30, 2026.” After that date, the playground, catalog, API, and BYOK endpoints would no longer be available — and customers needing continued model access were pointed to Microsoft Foundry or GitHub Copilot instead.

Hold your own tools against that bar. When your AI vendor next changes something material, does it give you a dated cutoff or a vague one? A rehearsal or a surprise? A destination or a shrug? The GitHub case also settles a comforting myth: that bringing your own key insulates you. BYOK — bring your own key, an arrangement where you plug your own account credentials for a model lab directly into a tool — went down with the ship here. GitHub’s own documentation now records that BYOK in GitHub Models is no longer available to any customer as of July 30, 2026, folded into the same retirement. BYOK changes who pays the lab; it does not change who controls the surface. Teams running production automations against models should pair this with our one-page model-pinning policy — the engineering version of this same discipline.

Everything above would be worth knowing in the abstract. The Cursor decision is what that looks like in practice. In a blog post published on Friday, August 28, 2026, OpenAI set out its decision to wind down the agreement that supplies OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026, and no future models to be provided as the agreement winds down. That is roughly two and a half months of notice — announced, dated, and public, which places it closer to the GitHub end of the spectrum than to a silent swap.

Notice what this event is and is not, using section 02’s vocabulary. It is not a model swap — Cursor is not quietly changing what answers your prompts. It is a supply wind-down of one provider’s models within a multi-provider tool: per Cursor’s own documentation, its supported frontier providers are OpenAI, Anthropic, and Google, and only the OpenAI-supplied portion is affected by this announcement. No user’s code, chat history, or account is touched by the announcement itself. What changes, if the shutoff proceeds as proposed, is which models appear in the picker.

The workaround fine print is where the reseller mechanics get vivid. OpenAI’s own Help Center article on using its models in Cursor explains that you can bring your own OpenAI API key — but that a key covers only Cursor’s local Chat and Agent features. Per the same article, it does not apply to Cursor Tab and autocomplete, Auto model selection and routing, Cloud or Background Agents, Automations, the Cursor CLI — its command-line version — or the Cursor API and SDK, the interfaces other software plugs into. And a detail that trips up non-engineers constantly: OpenAI states that a ChatGPT subscription does not include API usage — a bring-your-own-key setup is billed through a separate, metered OpenAI API account, not through the ChatGPT plan you may already pay for.

The last step in the ten-minute check is not a vendor disclosure page at all: it is the export setting inside your own account. If the access layer under a tool ever does retire, what you can carry out is whatever the export gives you — so it pays to know, in advance, exactly what that is.

Anthropic documents its version plainly. Anthropic’s help center states that individual Claude users on Free, Pro, and Max plans can export their information and chat history from Settings > Privacy, delivered as a ZIP containing the full conversation history in JSON format — a raw, machine-readable text format — with a download link that expires after 24 hours. The export is not available from the iOS or Android apps, and on Team and Enterprise plans only the organization’s Primary Owner can run it. For ChatGPT, third-party how-to guides describe a similar flow — an emailed ZIP with the conversation history as JSON plus a browser-readable copy, a link that expires after roughly a day, and a process that can take a few days — though we did not verify those mechanics against OpenAI’s own documentation, so treat the specifics as guide-reported rather than vendor-stated.

Now the finding that makes this section matter: an export is a copy, not a move. The files these tools hand you are conversation dumps a person can read or archive — not a structured history another tool can ingest. We checked the documentation of exactly three vendors — OpenAI, Anthropic, and Notion — and found no import counterpart to any of their exports. Three vendors is three vendors; we will not stretch that into a claim about the whole industry. But it is consistent with an argument the independent commentator Nick Sawinyh made in March 2026: that these exports exist primarily to satisfy privacy regulations like GDPR and CCPA, not to let you take your history somewhere else — in his words, a product decision, not an AI limitation.

The practical move costs you ten minutes once: run a test export from every AI tool your business depends on, open what comes back, and note two things — the format, and how long the download link lives. A 24-hour link means an export is something you schedule, not something you assume. Do this while nothing is wrong, and the worst-case scenario in section 02’s right-hand column becomes an inconvenience instead of a loss.

Here is the whole post as a procedure. Run it once per tool, and again whenever a renewal, a price change, or an odd shift in output quality gives you a reason. The first three steps point at pages that genuinely exist for almost every serious vendor — three of the four disclosure surfaces above. The fourth points at a setting inside your own account.

And when the answer is “we do not say”? Sometimes every step of the check comes up empty — no vendors named, no changelog, no public subprocessor page, no export. That is not neutral; it is a finding. Treat it three ways. First, ask in writing: subprocessor disclosure is a routine contractual request that any vendor selling to businesses has fielded before, and the quality of the answer tells you as much as its content. Second, weight the silence in your renewal decision — our 50-point vendor-risk template treats model provenance as one check among many, and our stack-audit framework covers when an opaque tool should be replaced outright. Third, remember the floor this post established: even at an opaque vendor, a model swap almost certainly does not touch your stored data — the risk you are pricing is discontinuity and quality drift, not deletion. The nastiest version of drift is when even the model’s name stays the same while the thing behind it changes — see what happens when a floating alias changes with no notice at all. If you would rather have this whole review run for you — supply chain, contracts, and replaceability across your AI stack — it is the kind of engagement our AI transformation practice exists for.