In October and November 2026, four announcements landed. OpenAI gave developers three months’ notice on 14 October that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the API on 14 January 2027, half its stated six-month commitment, with self-serve fine-tuning ending on 6 January. GitHub retired six Copilot models in one day on 19 October. Anthropic’s Opus 5.5 shipped four 400-error breaking changes, including a silent default-effort drop from high to medium and thinking blocks bound to the producing model. Cursor moved its cheapest tier to Grok-only models.
Together those four stories are one signal: model retirement became a cost your business can measure as an architectural property before any renewal. The three months OpenAI delivered against the six it promised is the say-do gap, the lens this page applies throughout. This guide maps where that cost lives and how to protect it. Portability is a property your team builds, and a vendor’s contract cannot grant it to you.
In This Series
- Model Deprecation Versus Silent Drift and What Breaks on a Swap: the two failure modes and what a swap breaks beyond the 404.
- Model Deprecation Notice Periods Across OpenAI Anthropic Google and Cloud Marketplaces: the actual notice windows and why the same model dies on different dates by channel.
- Where AI Exit Cost Accrues and What to Demand in Contracts: where switching cost concentrates and the contract asks that buy runway.
- Open Weight Models and Gateways Relocate Exit Cost Rather Than Remove It: what open weights and gateways move, and what they leave behind.
What does a model retirement actually look like — a loud 404 or silent drift under the same name?
Model deprecation is the announced, dated end of a named model: the endpoint keeps serving for a set window, then returns a 404 model_not_found. The riskier failure is silent drift, a latest alias that keeps returning 200s while the behaviour underneath changes. A dated snapshot fails loudly and on schedule, because the endpoint dies on a published date. An alias fails quietly and arbitrarily, because the same name keeps answering. The loud failure is easy to catch; drift is expensive because no error fires and only a regression suite detects it.
OpenAI’s notice is the loud case; Anthropic’s Opus 5.5 is the drift counterpoint, where token counts, fine-tunes and thinking blocks break without an error. Pinning makes swaps explicit, but the pinned model is the one that gets retired. Neither removes the problem, which is why what notice periods vendors actually give matters. A regression suite with a golden set, run in tools like Braintrust, Langfuse or Promptfoo, turns a swap into a measured verdict and is the same asset a fallback chain reuses.
Model Deprecation Versus Silent Drift and What Breaks on a Swap.
What are model deprecation notice periods actually worth across providers and cloud marketplaces?
A notice period is only worth what remains after you subtract your re-certification and change-control cycle. OpenAI’s 14 October 2026 notice delivered roughly three months against a stated six-month commitment, the say-do gap that defines the field. The same model retires on different dates depending on whether you call it through the first-party API, Amazon Bedrock or Microsoft Foundry, because each marketplace layers its own lifecycle policy over the vendor’s schedule. A notice that expires before your change-control cycle finishes is useless.
Anthropic’s 60-day commitment lives only in a documentation page and never appears in its commercial terms, while Google promises “advance notice” without a number. Bedrock runs a 12-month lifecycle with a 6-month Legacy period; Foundry fixes a retirement date at launch, up to 18 months out. Schedules live in official deprecation pages, Bedrock’s modelLifecycle field, Foundry’s lifecycleStatus field and machine-readable feeds. Runway only helps if you turn it into contract language, and self-hosting open-weight models is a third channel.
Model Deprecation Notice Periods Across OpenAI Anthropic Google and Cloud Marketplaces.
Where does AI exit cost actually accrue, and how do you protect your exit runway?
Exit cost accumulates in the agent layer around the model. That layer is the prompt schemas, tool definitions, memory, evals, guardrails and fine-tunes you have built on top of the endpoint, and each has to be re-validated when the model changes. The model endpoint is the cheap thing to swap; fine-tuning is the most expensive asset to move, because a fine-tune cannot migrate without retraining. The protection is contract language, meaning written notice-period commitments, disclosure of the model lifecycle, and export rights for your fine-tunes, negotiated before you sign the renewal.
An IBM survey found 71% of senior executives say switching their primary AI vendor would be difficult. That difficulty is the agent-layer cost now being priced. An honest migration-time estimate measured against what the actual notice periods are shows whether the runway you are offered is long enough. The remaining question is whether gateways and open-weight fallbacks actually remove the cost.
Where AI Exit Cost Accrues and What to Demand in Contracts.
Do open-weight models and gateways remove exit cost, or just relocate it?
Neither removes exit cost; both relocate it. Open-weight self-hosting means the vendor cannot take the model away, but serving those weights is your burden. A gateway moves the dependency from the model provider to the gateway and its assumed capability contract, the set of behaviours your workload depends on: tool calling, thinking blocks, context length and latency. If the gateway cannot tell a capability error from a transport error, failover produces silent regressions. While everything else stays behind, your regression suite is the one asset that travels.
For a closed model, retirement is a date on the vendor’s calendar; for open weights it becomes an operations problem, because you keep running stale weights and own the security and refresh burden. Failover must trigger on the right error class: a 404 advances the chain, while a 400 raises to the operator because every candidate will reject the same payload. Model IDs as runtime config plus an exercised cold-standby fallback is the pragmatic middle; candidates such as Meta, Qwen, DeepSeek, Mistral and Fireworks are judged against the capability contract. Open weights already handle 71.5% of Vercel AI Gateway tokens, and Eclipse’s Sovereign AI Foundation is an attempt to define the exit. The gateway debate needs where exit cost actually accrues and dated snapshots versus latest aliases.
Open Weight Models and Gateways Relocate Exit Cost Rather Than Remove It.
Resource Hub: Model Deprecation and Exit Cost Deep Dives
The Failure and the Calendar
- Model Deprecation Versus Silent Drift and What Breaks on a Swap: how a retirement announces itself, the loud 404 versus silent drift, and what a swap breaks beyond the endpoint.
- Model Deprecation Notice Periods Across OpenAI Anthropic Google and Cloud Marketplaces: the real notice windows across OpenAI, Anthropic, Google and the cloud marketplaces, and why the same model dies on different dates by channel.
The Cost and the Escape Hatch
- Where AI Exit Cost Accrues and What to Demand in Contracts: where switching cost actually accumulates, the agent layer rather than the endpoint, and the contract terms that determine who controls the exit.
- Open Weight Models and Gateways Relocate Exit Cost Rather Than Remove It: why pinning and gateways relocate rather than remove exit cost, and what open weights and fallbacks actually move.
Work the map in order: the explainer, the calendar, the contract, then the escape hatch, before your next renewal. If you have just seen a deprecation notice or are unsure what one looks like, start with the explainer. If you are mid-renewal or building a procurement position, start with the contract analysis. If you are deciding between channels or vendors, start with the calendar comparison. If you are evaluating gateways or self-hosting, start with the escape-hatch analysis.
The cost accrues long before renewal, in prompts, evals, fine-tunes and gateway semantics. Start with the article that matches where your exit currently hurts.
Frequently Asked Questions
What should I do when a vendor announces a model deprecation?
Lock down the exact date and the channel first, because the same model retires on different dates through first-party APIs, Bedrock or Foundry. Then inventory every workload still calling that model, pin the ones you are not migrating, and run your regression suite against any replacement candidate. Treat the notice as a running clock, not the day you start planning.
How long should I budget for re-certifying a replacement model?
Budget for the full re-certification and change-control cycle, not the swap itself. The endpoint change is the cheap part; re-validating prompt schemas, tools, evals and guardrails is where the weeks go, and a fine-tune adds retraining time on top. Subtract that total from the vendor’s notice period, and the remainder is the runway you actually have.
Can I move my fine-tuned model to a different provider?
Not directly. A fine-tune is trained against a specific model, so it cannot migrate to another provider; it has to be retrained from scratch. That makes fine-tuning the most expensive asset to move in your stack. If you depend on fine-tunes, negotiate fine-tune export rights and retraining runway before renewal, not after the deprecation notice arrives.
What does a “Legacy period” mean on Amazon Bedrock?
It is the tail end of Bedrock’s 12-month model lifecycle, a six-month window where a retired model still serves requests but is marked for removal. You get time to migrate, but you are running on a model that is already going away. Runway you cannot use is not runway, so start migrating when the Legacy flag appears, not when it ends.
Is self-hosting open-weight models cheaper than paying a vendor?
Not necessarily. You drop the per-token licence cost and the vendor’s ability to retire the model, but you inherit the operations: serving the weights, patching security and deciding when a stale model needs refresh. The cost relocates rather than disappears. Open weights change the shape of your exit cost, not its size.
Do I need a gateway if I only use one provider?
With one provider, a gateway is mostly a future option rather than a shield. It earns its place only if it can tell a capability error from a transport error, because failover should advance on a 404 and stop on a 400. If it cannot, you get silent regressions instead of clean fallbacks. The asset that actually travels is your regression suite, not the gateway.
What is the difference between a 404 and a 400 error during failover?
A 404 means the model no longer exists, so the fallback chain should advance to the next model. A 400 means the request reached a model but something broke, a capability or input problem, so the chain should stop and raise. Confusing the two is what turns gateway failover into silent regressions. The error class decides whether you fail over or stop.
How do I safely test a replacement model before switching?
Run a regression suite against a golden set of known inputs and expected outputs, then compare the candidate model’s verdicts against it. The golden set turns a swap into a measured verdict rather than a hunch, and it is the one asset that travels across providers. It also catches silent drift, where no error fires and only the suite notices.
Why do most buyers rank portability last at signature?
Because lock-in costs arrive later, while negotiating portability costs something now. Buyers value lock-in avoidance in principle, but at signature the renewal discount and the product roadmap win the room. By the time exit cost is real, the contract is closed. The fix is to price and demand the terms before renewal, while leverage still exists.
Is anyone building a standard for model deprecation?
Early efforts are underway. Eclipse’s Sovereign AI Foundation launched with 17 founding organisations to standardise how model exit is handled, and more than 70% of Vercel AI Gateway tokens now flow to open-weight models. It is still early, and a standard matters less than your own contract terms and regression suite, but the pressure to formalise exit is building.