When an AI Lab Cuts Off a Coding Tool: Cases Since 2025

Since June 2025, a frontier lab withdrawing or restricting model access affecting a named coding tool or peer lab has gone from unthinkable to documented fact — four times, by our count of the dated instances we could confirm. Anthropic limited Windsurf’s direct Claude access, revoked OpenAI’s API access, and technically blocked subscription OAuth tokens used outside Claude Code. Now OpenAI has notified SpaceX, Cursor’s new owner, that it intends to wind down Cursor’s access to OpenAI models.

If your team ships software with an AI coding tool — Cursor, Windsurf, Claude Code, Copilot, or one of the smaller harnesses — these events are your supply chain. The model behind your tool is provisioned under a contract between two other companies, and that contract can end on a schedule you do not control. What happened to Windsurf’s users in June 2025 took effect in under five days, by Windsurf’s own account.

This page is a ledger, not a narrative. It logs the dated instances with the action taken, the notice given, and the outcome. It also does two things most coverage skips: it separates the one peer-lab dispute from the downstream-tool cutoffs so the rows cannot be conflated, and it quotes — verbatim — what five major labs’ published terms actually say about assignment, change of control, and competing products. Candidates that did not survive verification are listed too, with the reasons they failed.

First, definitions. A supply cutoff, as this ledger uses the term, is a lab-initiated withdrawal or restriction of model access — not a tool switching vendors by choice, and not an outage. The rows split into two labelled groups because the events are not the same shape. In the first group, a lab restricts a downstream tool — a product that integrates the lab’s models and serves them to its own users. In the second, a lab restricts a peer lab’s own internal API usage — a competitor as a customer, not a reseller. Citing a row from one group as if it belonged to the other is the most common error in coverage of these events.

One more boundary worth naming: a lab retiring its own model IDs on a published schedule is a related but separate discipline — that calendar lives in our model deprecation and API sunset guide. This ledger only logs a lab restricting someone else’s access.

Aggregator coverage tends to compress these events into one headline — a lab “bans” a tool. The record shows four distinct mechanisms, and the differences are what a procurement or engineering lead actually needs to plan around. Three of the four were Anthropic actions, taken between June 2025 and February 2026, and each used a different lever than the one before it.

The pattern worth registering: only one of the four levers — the July 2025 revocation — was a publicly published contract clause being enforced as written. One was supply allocation with no clause cited, one was purely technical, and the newest one lives in a custom agreement the public has never seen. Reading a lab’s published terms tells you some of your exposure, not the whole of it — which is exactly why the second table in this post quotes what the published documents do and do not contain.

The first confirmed instance in this ledger. In early June 2025, Anthropic limited Windsurf’s direct first-party access to Claude 3.7 Sonnet and Claude 3.5 Sonnet — TechCrunch described the move as cutting off nearly all of Windsurf’s direct capacity. Windsurf said it received less than five days’ notice, a figure worth treating as vendor-stated since Anthropic never published its own timeline. Windsurf had already been denied direct access to Claude 4 at that model’s May 2025 launch, forcing a bring-your-own-key workaround for its users.

The context was an acquisition. OpenAI was reportedly in talks to buy Windsurf for roughly $3 billion, and Anthropic’s co-founder and Chief Science Officer Jared Kaplan addressed the decision directly — framing it as capacity triage toward customers with a future, with the acquisition rumor as subtext. Windsurf CEO Varun Mohan said publicly: “We have been very clear to Anthropic that this is not our desire — we wanted to pay them for the full capacity.”

The aftermath reshaped the company. The reported ~$3 billion OpenAI acquisition collapsed around July 11, 2025 — reportedly over Microsoft’s IP-access rights as an OpenAI investor. Within days, Google licensed Windsurf’s technology for $2.4 billion and hired CEO Varun Mohan, co-founder Douglas Chen, and other engineers; Cognition AI, the maker of Devin, then acquired the remaining company, announced July 14, 2025, as reported by CNBC. A supply restriction taken during deal talks preceded — no stronger causal claim than that — one of the stranger corporate unwinding sequences in the AI tooling market to date.

On July 29, 2025, Anthropic revoked OpenAI’s organization-level access to the Claude API. Before the mechanism, be clear about who this affected: no end users lost a tool. OpenAI was not reselling Claude inside a product — its own technical staff were using Claude Code internally, reportedly ahead of GPT-5’s launch. That makes this the ledger’s one peer-lab row: a supplier restricting a competitor’s direct API usage, a fundamentally different relationship from Windsurf, OpenCode, or Cursor.

Anthropic spokesperson Christopher Nulty’s statement, as reported by Wired and carried by PYMNTS, said Claude Code had “become the go-to choice for coders everywhere,” that OpenAI’s technical staff were using Anthropic’s coding tools ahead of the GPT-5 launch, and that this was “a direct violation of our terms of service.” The operative clause is quoted in full in the terms table below — Anthropic’s use restrictions bar accessing the services to build a competing product, including training competing models.

Two details keep this row honest. First, the revocation carried an explicit exception: Anthropic said it would keep providing OpenAI API access “for the purposes of benchmarking and safety evaluations as is standard practice across the industry” — so “total ban” framings are wrong. Second, OpenAI’s Chief Communications Officer Hannah Wong pointed at the asymmetry: “While we respect Anthropic’s decision to cut off our API access, it’s disappointing considering our API remains available to them.” GPT-5 launched shortly afterward as scheduled.

The third Anthropic action is the most misreported, so state what changed and who it affected first. On January 9, 2026, Anthropic deployed a server-side technical block on Claude Free, Pro, and Max subscription OAuth tokens being used outside the official Claude Code CLI and Claude.ai. If you paid for a Claude subscription and pointed a third-party coding harness at it, that stopped working. If you used a Claude API key — the metered developer path — nothing changed. “Anthropic banned Claude API access” is the wrong sentence; the API key path was never touched by this action.

The named affected tool was OpenCode, an open-source coding agent, whose users began seeing an error: “This credential is only authorized for use with Claude Code and cannot be used for other API requests.” The block affected third-party harnesses broadly that authenticated by presenting themselves as the Claude Code client. Per The Register’s reporting, Anthropic engineer Thariq Shihipar said the restriction addressed “unusual traffic patterns” that made it hard for Anthropic to help subscribers debug rate-limit and account issues.

The policy language followed the technical block by about six weeks. Around February 19–20, 2026, Anthropic updated its documentation to state — as reported by The Register — that using OAuth tokens from subscription plans in any other product, tool, or service, including the Agent SDK, is not permitted. We attribute that wording to The Register’s account rather than quoting it as Anthropic’s verbatim current text, since the underlying docs page wording can shift. Anthropic reportedly also sent legal requests, and OpenCode’s maintainers merged a change removing the Anthropic OAuth plugin, reported mid-March 2026.

For teams, this row created a distinction that did not exist before: subscription-auth integrations and API-key integrations now carry different risk profiles. The first runs on consumer terms a lab can technically enforce overnight; the second runs on commercial terms with (some) published assignment and restriction language you can actually read.

The newest row, and the reason this ledger exists now. On August 28, 2026, OpenAI published a blog post stating it intends to wind down the contract that supplies OpenAI models to Cursor, with a proposed shutoff date of November 12, 2026. The trigger: SpaceX’s $60 billion all-stock acquisition of Anysphere, Cursor’s maker, which closed August 14, 2026. Precision matters here more than anywhere else in the ledger — nothing has been cut off as of August 28, 2026. This is a notification of intent with a 76-day window, which OpenAI describes as “the maximum notice provided by our contract.”

A change of control — the operative concept — is a contract event where a party’s ownership changes hands, and some agreements let the counterparty renegotiate or cancel when it happens. OpenAI’s stated reason leans on history with the acquirer: “We are making this choice because we 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.” The post cites Twitter breaking contract terms after Musk’s acquisition and Musk’s admission under oath earlier in 2026 that xAI had violated OpenAI’s terms of service.

That sentence contains this post’s most useful finding, unpacked in the terms table below: OpenAI’s published Business Terms never use the phrase “change of control.” The phrase appears only in OpenAI’s characterization of its custom, non-public Cursor contract. What the deal-makers signed and what the public terms say are different documents — and if your own vendor contract is custom, the published terms tell you less than you think.

The event itself is covered in depth in our companion piece on OpenAI’s decision to end Cursor’s model access; the acquisition backdrop is in our coverage of the SpaceX–Anysphere close. This ledger holds the row itself: notified August 28, proposed shutoff November 12, outcome open.

A ledger is only as trustworthy as its rejections. Six candidate events came up during research that read, at headline level, like they belonged here — and did not survive verification against the class definition: a dated, confirmed instance of a frontier lab withdrawing or restricting model access affecting a named coding tool or peer lab. Listing them is evidence, not an appendix; it shows where the boundary of the class actually sits.

One neighboring event class is worth flagging because it gets conflated constantly: outages. When GitHub’s August 2026 incident took Copilot down with it, developers lost their coding tool for hours — but no supplier withdrew anything. That failure mode and its dependency chain are covered in our analysis of the GitHub–Copilot outage. Outages end when the incident ends; the rows in this ledger are decisions.

Most coverage of the Cursor decision treated “change of control” as standard boilerplate. So we pulled the operative documents — each lab’s currently published commercial terms or license — and quoted the assignment and competitor-use language verbatim. Where a clause was not located in the documents reviewed, the cell says so rather than summarizing something unread. Two structural notes first: Meta’s entry is a model-weights license, not an API terms of service, a different kind of instrument from the other four; and a published document is not the same thing as the custom contract a large customer may actually sign.

Three things fall out of reading the actual text. First, the two labs that have acted most — OpenAI and Anthropic — both key assignment to a “successor to substantially all” assets or business, not to “change of control”; that phrase exists, on the public record, only in OpenAI’s description of its custom Cursor contract. Second, competitor-use restrictions vary more than the coverage implies: OpenAI and Google restrict competing-model development, Anthropic adds an explicit resale bar, xAI’s enterprise terms carry a training restriction instead, and Meta’s gate is a user threshold on a weights license. Third, an empty cell is information: no explicit assignment clause was located in the Google documents reviewed, and inferring one would be exactly the kind of summarizing this table exists to avoid.

The four rows do not line up into a single direction of travel — each instance used a different lever than the one before it. One row, Cursor in August 2026, was triggered by a completed ownership change; the other three were not, so treat that as one documented case rather than an established pattern. The notice you get is whatever the paper says: in these instances, from under five days (vendor-stated) to 76 days (a stated contract maximum).

Looking forward, the reasonable projection is not that cutoffs become common — four dated instances in fifteen months is a low base rate — but that the exposure concentrates where relationships are custom and ownership is in motion. If your coding tool is owned by, or being acquired by, a company that competes with its model supplier, the Cursor row is now the template for how that ends: a blog post, a notice period, and a date. Plan for the date, not the blog post.

The practical toolkit already exists. Use the 50-point AI vendor risk assessment to score the relationship before it becomes a ledger row, and the cross-harness skill portability analysis to measure how movable your setup actually is. The procurement framing — why enterprise buyers should treat frontier labs as strategic suppliers rather than utilities — is developed in our enterprise vendor-risk read of the frontier labs. And if you want a second set of senior eyes on your own AI supply chain — which tools, which models, which contracts, which fallbacks — that is the kind of dependency mapping our AI transformation engagements start with.