Read the standard terms of the five places most enterprises buy frontier models and the time you get to pull your data out after termination runs from zero to 90 days. Microsoft gives 90, AWS gives 30, Google expects you to ask for your data back before the term ends, Anthropic's data processing addendum returns it if you ask within 30 days, and OpenAI deletes it within 30 days with a return right that covers only personal data. If you can buy models through a cloud agreement you have already negotiated, do that. Whatever paper you sign, put five exit riders on the order form, because none of the five standard agreements promises transition assistance.
Every term below was read on the vendor's live legal or documentation page on October 5, 2026. These are the click-through or standard versions, and a negotiated order form overrides every one of them.
The Verdict and the Comparison Table
The weakest exit paper of the five is Anthropic's Commercial Terms (effective June 17, 2025): data return only through the incorporated DPA and only if you ask within 30 days, a contractual right for Anthropic to assign the agreement in a sale of its business without your consent, and model retirement notice that lives in documentation instead of the contract. The strongest is the Microsoft paper you already have for Azure, which gives you a 90-day extraction window and an 18-month model runway. AWS and Google sit in between, each strong on one clause and weak on another.
| Exit term (standard paper, read 2026-10-05) | OpenAI | Anthropic | Microsoft Foundry | Google Vertex AI | Amazon Bedrock |
|---|---|---|---|---|---|
| Data out after termination | Deleted within 30 days; DPA return right covers personal data only | Returned if requested within 30 days (DPA) | 90 days to extract from a disabled account | Ask during the term; deleted within 180 days after a 30-day recovery period | 30 days to retrieve, if fees are paid |
| Model retirement notice | 6 months for GA models (docs) | 60 days (docs) | GA models retire 18 months after launch, 60-day email | 12 months for a discontinued Service, unless replaced | Per model card; 6-month or 45-day Legacy period |
| If the vendor is sold | Either side may assign to a successor on 30 days' notice | Anthropic may assign; you need consent | Consent needed beyond affiliates | Consent needed beyond affiliates; 30-day notice of change of control | AWS may assign in a merger; you need consent |
| Price change notice | 14 days after Pricing Page update | 30 days | Terms fixed for the subscription | GCP fees may change at any time unless the order form says otherwise | 30 days |
| Transition assistance | None; unpaid minimum commitment falls due | None | None | None | None beyond the 30-day window |
The defined workload behind every judgment on this page: a three-year commitment, a retrieval assistant over 5 million documents (about 10 billion tokens) indexed in the vendor's hosted vector store, one fine-tuned model, and a minimum annual spend in the low millions. That is the profile where leaving costs real money and where the gaps above turn into months.
Clause One: Can You Get Your Data and Embeddings Out?
Data export on termination is the clause that decides whether leaving is an engineering project or a restart. Read with their data processing addenda, three of the five standard agreements give you a defined window after the contract ends.
The AWS Customer Agreement (updated August 14, 2026) says that for 30 days after termination AWS will not remove your content "as a result of the termination" and will let you retrieve it, but "only if you have paid all amounts due." Microsoft's Product Terms for Online Services give "90 days to extract Customer Data from a disabled account." Google's Cloud Data Processing Addendum puts the burden on you: if you want data after the term, you instruct Google to return it during the term; afterwards Google deletes it after "a recovery period of up to 30 days" and "within a maximum period of 180 days."
The two model labs put their answer in the addendum, not the main agreement. Anthropic's Section I.4 says only that "Customer may no longer access the Services," but its Section C incorporates the Anthropic Data Processing Addendum, whose Section H returns "a copy of all Customer Data" within 30 days of termination "if requested to do so by Customer within that period," then deletes it. Miss the window and the data is gone. Section 11.3 of the OpenAI Services Agreement (online version v.010126) says that on termination "OpenAI will delete all Customer Content from its systems within thirty days." The OpenAI DPA (v.010126) adds that OpenAI will "return or delete Customer Data" at your instruction, but it defines Customer Data as personal data, so your non-personal documents and outputs have no stated retrieval right.
Embeddings are where legal teams usually over-negotiate. A vector is only useful alongside the model that produced it. When Google retired text-embedding-004 from the Gemini API, one open-source project's migration ticket (opened February 6, 2026) recorded the cost: the new model emits 3,072 dimensions against 768, so "all indices must be deleted and re-created." A developer forum thread from October 2025 shows users getting migration emails while the documentation still listed the model as stable.
Re-embedding itself is cheap. At OpenAI's list rates derived from its embeddings guide (62,500 pages per dollar for text-embedding-3-small and 9,615 for text-embedding-3-large, at roughly 800 tokens a page, checked October 5, 2026), our 10 billion tokens costs about $200 or $1,300 to re-embed. The export worth fighting for is your chunked text, chunk boundaries, metadata and retrieval evaluation set, in a documented format, because those take weeks to rebuild and the vectors take an afternoon.
Clause Two: How Long Does the Model You Tested Stay Alive?
Model continuity is the commitment that the exact version you validated keeps answering until you choose to move. Three of the five vendors put it in documentation, which they can edit, and two put something close to it in the contract.
OpenAI's deprecation page promises "at least 6 months" for generally available models, 3 months for specialized variants and as little as 2 weeks for previews, and less "if safety or compliance concerns require us to retire a model sooner." Anthropic's deprecation page promises "at least 60 days' notice before model retirement for publicly released models." Its dated history shows the pattern: Claude Sonnet 4.5 was deprecated on September 30, 2026 and retires November 30, 2026. Anthropic has separately committed (November 4, 2025) to preserve the weights of publicly released models for the life of the company, which helps researchers and does nothing for your production endpoint.
Microsoft's Foundry lifecycle policy is the most predictable: GA models get a retirement date "18 months out" at launch, at least 60 days of email notice, and the answer to "Can I get an exception to extend a model's retirement date?" is "No." Claude, DeepSeek, Fireworks and Mistral models on Foundry follow a 12-month lifecycle instead. The retirement schedule also shows fine-tuned gpt-4o deployments retiring on 2027-10-01, which is the date your tuned model stops answering regardless of what you paid to train it.
Amazon Bedrock's lifecycle page prints an "EOL no sooner than" date on each model card and a Legacy period of 6 months or 45 days. Watch the exception: GPT-6.1 Sol on Bedrock "follows the OpenAI first-party model's lifecycle" and has "no separate Bedrock minimum-availability date," so OpenAI models bought through AWS stay on OpenAI's notice periods. We compared the notice windows platform by platform in our deprecation policy breakdown.
Google's contract term is the strongest on paper. Section 1.4(e) of the Google Cloud Terms of Service requires 12 months' notice before discontinuing a Service "unless Google replaces such discontinued Service or functionality with a materially similar Service or functionality." A newer model of the same family is easy to call materially similar. Define the model version as the Service in your order form or that 12 months may never start.
Clause Three: What Happens When Someone Buys the Vendor, or You?
The change-of-control clause decides who your counterparty is after an acquisition. On standard paper, three of the five vendors can hand your contract to a buyer without asking you.
OpenAI's Section 16.7 lets either party assign the agreement "to a successor to substantially all the respective party's assets or business" with 30 days' written notice. Anthropic's Section M.4 lets Anthropic assign "as part of a sale of all or substantially all its business" and requires your consent for anything you do. AWS's Section 11.1 is the same shape: AWS may assign "in connection with a merger, acquisition or sale," and you need consent. Google requires written consent for assignment outside affiliates in both directions and obliges a party to give notice "within 30 days after the change of Control." The Microsoft Customer Agreement also requires the other party's written approval for any assignment beyond affiliates.
A notice period gives customers very little when the buyer wants the team and not the product. When HP bought Humane's software and patents for $116 million, it left the AI Pin out of the deal; the devices stopped working ten days after the announcement and stored customer data was deleted. Acquisitions also work in reverse. When reports surfaced that OpenAI was buying Windsurf, Anthropic cut most of Windsurf's direct Claude access with less than five days' warning, according to Windsurf's CEO. Your own acquisition by a vendor's rival is a change-of-control event too, and nothing on standard paper protects you from it.
The clause to ask for is a termination right for you, without penalty and with the unused commitment refunded, if the vendor is acquired by a party you name or if you are. We have tracked how rarely acquirers say anything about the acquired product's customers, from Mitratech's BotDojo deal to Progress buying Domo's contracts.
Clause Four: Price Protection and the Commitment You Cannot Cancel
Price protection is the commitment that the per-token rate you modelled is the rate you pay for the term. The standard notice periods are short, and the commitments run the other way.
OpenAI's Section 6.6 makes "Price changes on the Pricing Page" effective "fourteen days after they are posted." Anthropic's Section H.1 gives 30 days. AWS's Section 3.1 lets it "increase or add new fees" on 30 days' notice. Google's Section 2.6 says it "may change the Fees at any time unless otherwise expressly agreed in an addendum or Order Form," with the 30-day notice limited to Workspace, Looker (original) and Cloud Identity. Microsoft's Product Terms say the terms in force when you renew or buy "will not change during Customer's subscription," which protects the terms more clearly than a metered rate, so put the rate card in the order.
The commitment is where leaving gets expensive. OpenAI's Section 6.1 makes a minimum commitment "non-cancellable," Section 11.3 makes "any unpaid minimum commitment amounts" due immediately on any termination except one you make for cause, and Section 1.2 lets OpenAI "adjust or remove discounts" if you reduce your commitment at renewal. Ask for two things: a cap on any rate increase at renewal, and a commitment that converts into spend on a successor model or rolls over when a model you depend on is retired. Our reserved versus spot GPU analysis makes the same case for compute: commit for the period you can forecast.
Clause Five: Assistance Period and Migration Support
An assistance period is a defined time after notice during which the vendor keeps the service running at the contracted price and helps you move. None of the five standard agreements has one.
Regulation is starting to write it for European buyers. The EU Data Act's switching rules, summarised by DLA Piper (July 7, 2026), cap the notice period at two months, set a 30-day transition period that can stretch to seven months only if the provider justifies it within 14 working days, give at least 30 days to retrieve exportable data, and ban switching charges, including egress, from 12 January 2027. The rules cover SaaS and PaaS as well as infrastructure. The Act does not mention embeddings or fine-tuned weights by name, so define "exportable data" yourself. We covered how the related DMA designations land mid-commit in our piece on AWS and Azure.
Builder.ai shows what happens with no assistance period at all. When it entered insolvency in May 2025, the platform froze and customers were left with half-finished apps and no support. A contract cannot fix insolvency, but an escrowed export, delivered on a schedule while the vendor is solvent, can.
Who Should Not Sign Each Vendor's Standard Paper
OpenAI direct is the wrong standard paper for a buyer signing a minimum commitment it might not consume, because unpaid commitment falls due on termination. It does carry a protection the others lack: under Section 2.3, if an update "materially reduces the Services functionality," you can terminate by giving notice within five business days. Get that clause extended to cover model retirements in writing.
Anthropic direct, on the standard Commercial Terms, should not carry anything with stored data or a multi-year commitment. You get a 30-day data return that lapses if you forget to ask, the shorter documented model notice of the two labs, and assignment that runs one way. Claude bought through Bedrock or Foundry sits under cloud paper instead, and our deprecation comparison found Bedrock keeping Claude about four months past Anthropic's own retirement date.
Microsoft Foundry is wrong for a workload that depends on one model past its 18-month life (12 months for Claude and other partner models), because retirement extensions are refused as policy. Provisioned deployments are not auto-upgraded either, so the migration work stays with you.
Google Vertex AI is wrong for a buyer who cannot get fees written into the order form, since the standard terms let GCP fees change at any time. It also punishes a team that forgets to request data return before the term ends.
Amazon Bedrock is the wrong place to buy OpenAI models for AWS's lifecycle guarantees, because those models follow OpenAI's. Expect a fight at exit if a billing dispute is open, since retrieval depends on paying "all amounts due."
How to Decide: The Criteria That Predict Regret
Choose the paper by what you will need on the day you leave. Four questions predict regret better than the price per million tokens.
- Where does your retrievable state live? If chunks, metadata and evaluation sets sit in a vendor-hosted store, the data-out clause outranks everything else. If they sit in your own database, it barely matters.
- How long is your validation cycle? If re-certifying a model takes a regulated team four months, Anthropic's 60 days and Bedrock's 45-day Legacy period are shorter than your process.
- Could your vendor, or your company, be acquired during the term? If yes, a termination right on change of control is worth more to you than a discount.
- What happens to the unused commitment if you terminate early? Model that case before you model the renewal.
A negotiated order form changes every row of the table. OpenAI's Section 16.2 ranks the Order Form above the agreement itself, so the weaker a vendor's standard terms, the more of the riders below you need in that form.
Five Riders to Put on the Order Form
This Week: Pull every AI agreement you have signed and fill in the table above with your actual terms, not the vendor's standard ones. Flag any row that says "not addressed."
This Month: Draft the riders. (1) Export of chunked content, metadata, prompts, evaluation sets and fine-tuning data in a documented format, on request during the term and for 90 days after it. (2) Twelve months' notice before retiring any model version named in the order, with the commitment convertible to its successor. (3) A no-penalty termination right with a refund of unused commitment if either party undergoes a change of control. (4) A rate cap at renewal and no mid-term rate increase. (5) A 90-day assistance period at contracted rates after any notice of termination.
Before Signature: Ask the vendor to show where each rider appears in the order form itself. OpenAI's Section 16.1 says the agreement "supersedes all prior or contemporaneous agreements, communications and understandings," so a promise made in email or on a sales call has no force once you sign.
Continue Reading
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic
- AWS and Azure DMA Duties Land Midway Through a 3-Year Commit
- OpenAI Agents SDK Alternatives: Leave Three Features, Not the SDK
- Mitratech Buys BotDojo, Whose Terms Offer No Data Export
- Reserved vs Spot GPUs: Commit Six Months, Not Twelve
- AI Data Portability War: What It Means for Your Vendor Strategy
