Buyers of AI voice and sales agents interrogate accuracy, integrations and data handling. Almost nobody asks the question that decides whether the agent still works in eighteen months: what happens when the model underneath it is switched off?
Quick answer
Foundation models are retired on published schedules, and a retired model stops serving requests entirely. “Deprecated” means still working but on the clock; “retired” or “shut down” means calls fail. Fine-tuned models are pinned to their base model’s lifecycle. Before signing with an agent vendor, get in writing who owns the migration, what notice you receive, and who pays to revalidate the agent afterwards.
The vocabulary, in the providers’ own words
These four words are not synonyms, and confusing them is how a team ends up with two weeks’ warning instead of six months. Every quote below was read from the provider’s own documentation on 30 August 2026.
Deprecated. OpenAI: “We use the term ‘deprecation’ to refer to the process of retiring a model or endpoint. When we announce that a model or endpoint is being deprecated, it immediately becomes deprecated.” Anthropic calls a deprecated model “still functional but no longer recommended”. It keeps working — for now.
Legacy. The signal that usually precedes deprecation. OpenAI uses it “to refer to models and endpoints that no longer receive updates”. Amazon Bedrock makes it a formal state with a clock attached: “A model will be in the Legacy state for at least 6 months before the EOL date.”
Retired, shut down, sunset, EOL. The point of no return. Anthropic: “Requests to retired models will fail.” Microsoft is most precise about the failure mode — a retired model is “Removed from service. All inference requests return 410 Gone.”
Snapshot pinning. You can call a moving alias or a dated version. Microsoft defines a model version as “a dated release within a family (for example, gpt-4o 2024-05-13 vs. 2024-08-06)”. Pinning buys stability and costs you an expiry date; an alias buys an evergreen endpoint and costs you control over when the behaviour changes.
What actually breaks
Fine-tunes die with their base model
This is the one that surprises people who paid for a custom model. OpenAI’s deprecations page states it plainly: “Inference on fine-tuned models will be disabled only when the underlying base model is deprecated.” That reads as reassurance until you open the tables — OpenAI’s 22 April 2026 legacy-snapshot notice lists fine-tuned versions with their own shutdown date of 23 October 2026, including ft-gpt-4 and ft-gpt-3.5-turbo. Microsoft runs two clocks (“Fine-tuned models retire in two phases: training and deployment”), and on Bedrock, once a base model goes Legacy, “you cannot create new fine-tuning jobs on that model.”
If your agent’s tone and qualification logic live in fine-tuned weights rather than portable prompts, a base-model retirement is a rebuild — one reason the choice between fine-tuning, prompt engineering and retrieval for AI sales agents deserves a decision on the record.
Voice and transcription models have their own calendar
A phone agent is a stack — speech-to-text, a language model, text-to-speech — each with its own lifecycle. OpenAI’s deprecations page carries a 20 July 2026 notice covering “legacy audio, realtime, and transcription model families and snapshots of their deprecation and removal from the API on January 20, 2027”, listing gpt-realtime, gpt-4o-realtime and older transcription snapshots. A text-to-speech change alters how your agent sounds — pacing, prosody, interruption handling — without a word of your script changing, which makes it a brand problem as well as an engineering one.
Behaviour, cost and latency all move
A replacement is not a drop-in. Microsoft tells customers to “Compare quality, latency, and cost together rather than relying on public benchmarks alone.” Cost can move against you: Bedrock warns that during a Legacy model’s extended-access phase “you should expect higher pricing, which will be set by the model provider.”
The same model can have different dates depending where you buy it
Anthropic states that “Partner-operated platforms (Amazon Bedrock and Google Cloud) set their own retirement schedules, so a model’s lifecycle status and dates can differ.” Bedrock says the same from the other side: its dates “may differ from dates published by model providers (such as Anthropic or Cohere).” So “we run Claude” tells you nothing about your retirement date. The channel does.
The published schedules — who says what
| Policy statement | Owner (organisation) | Source URL | Date checked |
|---|---|---|---|
| Minimum notice: GA models “At least 6 months”; preview models “may be retired with much shorter notice, such as 2 weeks” | OpenAI | developers.openai.com/api/docs/deprecations | 30 Aug 2026 |
| “at least 60 days’ notice before model retirement for publicly released models”; Active / Legacy / Deprecated / Retired lifecycle | Anthropic | platform.claude.com/docs/en/about-claude/model-deprecations | 30 Aug 2026 |
| Per-model dates; retirement timelines “may be extended” but “won’t be moved to an earlier date than what is listed”; short-term availability models “retire 45 days after a replacement model is released”. This page states no general advance-notice period. | Google Cloud | docs.cloud.google.com — Model versions and lifecycle | 30 Aug 2026 |
| GA notice “At least 60 days”; preview “At least 30 days”; dates “aren’t extendable”; provisioned deployments “are NOT auto-upgraded” | Microsoft | learn.microsoft.com — Foundry Models lifecycle and support policy | 30 Aug 2026 |
| Models remain “at least 12 months before the EOL date”; Legacy lasts “at least 6 months”; “Migration will not happen automatically.” | Amazon Web Services | docs.aws.amazon.com — Amazon Bedrock model lifecycle | 30 Aug 2026 |
Two things to carry into a vendor conversation. For generally available models the published minimum notice differs by a factor of three — six months at OpenAI against 60 days at Anthropic and Microsoft — and preview models get less again at both. Notice goes to the account or subscription holder. If that is your vendor, the clock starts without you.
Who owns the migration: three buying models compared
| Build on the raw provider API | Buy an agent platform | Buy a managed service | |
|---|---|---|---|
| Who decides the migration date | You, inside the provider’s window | The vendor, unless the contract says otherwise | The provider decides; the firm schedules the work |
| Notice you receive | Direct from the provider, on its published schedule | Whatever the contract says — check whether it says anything | Whatever the statement of work says |
| Who revalidates | Your team | Vendor for the platform; you for your scripts and outcomes | Billable to you unless the contract says otherwise |
| Revalidation cost | Internal engineering and QA time | Ask whether it is absorbed, and whether “material change” reopens pricing | Typically a change request |
| Portability of prompts and config | Total — you hold the files | Depends on export capability; demand a live export before signing | Depends who authored the assets and who owns the IP |
The contract questions to ask any agent vendor
- Which models, by exact identifier, does my agent call today? A family name is not an answer.
- Are you the account holder with the model provider, or am I? That decides who gets the retirement email.
- How much notice before you change the model behind my agent, and is that contractual or a courtesy?
- Is anything in my agent fine-tuned? On which base model, and what happens to those weights at its retirement date?
- Who pays for revalidation? Name the party in the contract.
- Do I get an acceptance test before a model swap reaches production traffic?
- Can I export my prompts, scripts, configuration and evaluation set today, in a format readable without your product? Ask for the export, not confirmation.
- What happens to recordings, transcripts and evaluation sets after a migration, and for how long?
- If the voice model changes, do I approve the new voice?
These belong in the same document as our voice AI vendor security questionnaire. A vendor who answers the security questions crisply then goes vague on lifecycle has told you something.
How to build so a model swap is boring
Keep a golden set. Fifty to two hundred representative conversations with known-good outcomes: qualification decisions, objection handling, compliance behaviour, escalation triggers. It tells you whether a replacement is better, worse or merely different — and you cannot build that baseline from a model already switched off.
Pin snapshots, then diarise the review. Pinning without a calendar entry is how teams learn a retirement date from an error log. Microsoft’s own advice: “Don’t wait for Microsoft to name an official replacement before you begin evaluating newer models.”
Treat a model swap as a release. Change log, staged rollout, rollback plan, sign-off — not a configuration tweak done on a Friday. Set auto-upgrade deliberately too: Microsoft documents a NoAutoUpgrade option with the consequence stated plainly — “deployment stops working at retirement”. Both choices are defensible; choosing by accident is not.
Where Zian sits, honestly
Zian AI does not pretrain or own a foundation model, and we will not imply otherwise. Our layer sits above it: SmartReach AI™ orchestrates message, channel and timing by country, industry and profile, and PrecisionPitch AI™ continuously split-tests scripts against real success outcomes — set out in our note on proprietary AI sales tech.
A layer that continuously split-tests against outcomes is already measuring performance rather than assuming a model version behaves as it did last quarter. That helps; it is not a guarantee. We publish no migration SLA, no notice period and no migration cost commitment. We also support private model deployment on customer infrastructure, which changes the shape of the problem without removing it.
Frequently asked questions
What happens to my fine-tuned model when the base model retires?
It goes with it. OpenAI’s deprecations documentation states that “Inference on fine-tuned models will be disabled only when the underlying base model is deprecated”, and its April 2026 legacy-snapshot notice lists specific fine-tuned versions with a 23 October 2026 shutdown date. Amazon Bedrock separately blocks new fine-tuning jobs once a base model enters the Legacy state. Budget for retraining, not a rename.
How much notice will I actually get?
It depends on the provider and the tier. OpenAI commits to “At least 6 months” for generally available models but warns that preview models “may be retired with much shorter notice, such as 2 weeks”. Anthropic commits to “at least 60 days’ notice before model retirement for publicly released models”, and Microsoft states at least 60 days for GA and 30 for preview. Google Cloud publishes per-model dates that “won’t be moved to an earlier date than what is listed”, but its model-versions page states no general advance-notice period. All of it goes to the account or subscription holder — check whether that is you or your vendor.
Do model version changes really change agent behaviour?
Measurably, yes. In the arXiv preprint “How is ChatGPT’s behavior changing over time?”, Lingjiao Chen, Matei Zaharia and James Zou compared the March 2023 and June 2023 versions of GPT-3.5 and GPT-4 and reported that “the behavior of the ‘same’ LLM service can change substantially in a relatively short amount of time” — GPT-4 identified prime vs composite numbers at “84% accuracy” in March and “51% accuracy” in June. That is version drift rather than retirement, but it is why a golden-set regression suite is not optional.
Who pays for revalidation after a forced migration?
No industry default exists, which is why it belongs in the contract. Model providers do not compensate anyone for migration effort: Microsoft confirms retirement dates “aren’t extendable”, and Amazon Bedrock states that “Migration will not happen automatically.” Between you and your vendor, decide whether revalidation is absorbed as maintenance or billed as a change request, and name it.
Does Zian publish a model migration SLA?
No. Zian AI is in partnership-application beta and publishes no migration SLA, no security certification and no pricing. We will answer these questions in writing during your application. If any vendor — including us — gives you a verbal assurance about model migration, ask for it on paper. What platforms do and do not commit to is set out in our AI sales agent platform capability matrix.
Ask the deprecation question early
The retirement schedules are public and dated. The gap is that the question rarely reaches the evaluation checklist until an agent starts returning errors. Ask every shortlisted vendor the same nine questions and compare the written answers.
If you would like ours in writing, Apply For Partnership and put model lifecycle in your first message.