The notice a vendor quotes is not the runway you get. OpenAI promises six months for generally available models, yet its 14 October 2026 notice gave GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini three months, with no explanation offered.
Anthropic has settled on its 60-day floor, Google publishes no numeric notice for stable models, and Amazon Bedrock and Microsoft Foundry add their own retirement dates to the model maker’s schedule. A notice period matters only when you compare it with your re-certification cycle, and that comparison is what determines the exit cost of your AI stack.
How much notice do OpenAI, Anthropic, and Google actually give before retiring a model?
OpenAI’s six-month promise slipped to three months in its October 2026 notice, Anthropic has converged on its 60-day floor, and Google publishes no numeric notice for stable models.
OpenAI’s policy is six months for general models and three for specialised variants, which a normalised comparison documents. Self-serve fine-tuning for these models ends even earlier, on 6 January 2027, eight days before the models themselves retire.
Anthropic walks each model through four lifecycle states, from Active while it’s the current model, to Legacy once a newer model has replaced it, to Deprecated when a retirement date is set but it still serves, and finally Retired when it stops serving for good. The last five batches of models Anthropic retired got 60, 61, 61, 62 and 62 days of notice, and the cohort arithmetic shows the real windows have shrunk from about 180 days in early 2025 down to the 60-day minimum. Anthropic has also changed a model’s behaviour without an official retirement: Opus 5.5 changed token counts by up to 1.35 times, the response shape, and sampling parameters in version 2026-10-21, what breaks on a swap in practice.
Google’s Gemini API publishes shutdown dates as “earliest possible” and promises notice without a number, giving two weeks only for previews and the latest alias. Vertex AI runs a separate calendar where dates can be extended but never moved earlier.
Why does the same model retire on different dates across the first-party API, Amazon Bedrock, and Microsoft Foundry?
Each marketplace layers its own lifecycle policy over the model maker’s schedule. Amazon Bedrock layers an Active/Legacy/EOL lifecycle over Anthropic’s dates and says only its own page’s dates apply. Microsoft Foundry runs an 18-month lifecycle, 12 months for Anthropic and DeepSeek models, and after retirement the endpoint returns a 410 Gone error to tell you the model is permanently gone.
Claude Sonnet 4 retired on Anthropic’s API on 15 June 2026 but stays on Bedrock until 14 October 2026. Claude Opus 4.1 retired on 5 August 2026 but runs on Bedrock until 8 January 2027. Gemini 2.5 is retired on Vertex AI while still not deprecated on the Gemini API, which the deprecation calendar tracks.
A marketplace can be a buffer or an extra failure point. Bedrock guarantees at least 12 months on-platform plus a Legacy period, though the Legacy period carries a 15-day inactivity clause and the extended-access phase that follows comes at higher pricing set by the model provider. Foundry says retirement dates are not extendable and auto-upgrades standard deployments unless you set NoAutoUpgrade, while Anthropic notes partner platforms set their own schedules.
How do you tell whether a vendor’s deprecation notice period is long enough for your re-certification process?
Compare the vendor’s window to your change-control duration. The notice window minus your re-certification time is your runway.
Re-certification is the cycle a swapped model has to pass before it goes live. Security reviews the new model, legal signs off on its terms, and your team runs it against your eval suite to catch regressions. If that takes a quarter, Anthropic’s 60 days and Google’s preview two weeks are both too short. A six-month notice is worthless if your change-control takes seven. That leaves you with negative runway.
Fine-tuning and provisioned deployments enlarge the cycle. A study of 22,555 migration commits sized a fine-tuning swap at roughly 700 lines across 14 files, against about six lines for a prompt-only app. Fine-tuned models follow their base model’s schedule, as the retirement rules lay out. The window only pays off once you turn it into contract language.
Should you buy frontier models through the first-party API or a cloud marketplace?
It depends on whether you value control or convenience. The first-party API gives earliest access, one deprecation policy, and one retirement calendar. A marketplace gives consolidated billing, governance, and sometimes a longer legacy window, at the cost of a second lifecycle layer.
Bedrock’s Claude Sonnet 4.5 went Legacy on 8 October 2026 and reaches end of life on 8 April 2027, four months after Anthropic retires the same model on 30 November 2026, which the published dates confirm.
Data residency can tip the call for your business. If your residency rules dictate where your data can land, the differences matter: OpenAI models on Azure offer EU data zones, while Claude on Foundry has no EU data zone, and the US CLOUD Act applies to Anthropic. That’s the compliance picture you weigh against your own requirements. If you self-host an open-weight model, the vendor-set retirement date disappears, and the serving, patching and capacity-planning work moves to you.
Why did GitHub retire six Copilot models in one day, and what does that signal about the AI coding assistant market?
GitHub cleared six superseded Copilot models at once and replaced them with GPT-6.1 Sol. For buyers, churn arrives as a batch rather than a schedule. The 19 October 2026 deprecation retired GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash and Grok 4.5, all replaced by GPT-6.1 Sol.
Cursor shows how ownership events become deprecation risk. After SpaceX acquired Anysphere, Cursor’s parent company, Cursor’s OpenAI model supply was proposed to wind down under a change-of-control clause, a roughly 76-day “proposed” window ending 12 November 2026.
Windsurf, the AI coding assistant, got under five days of notice when Anthropic limited its Claude access in 2025.
Cursor’s cheapest tier runs on Grok alone. Its bring-your-own-key option covers local Chat and Agent, and it does not cover Tab autocomplete, Auto model routing, or the API/SDK.
DeepInfra handles retirement differently. Where the other tools cut you off on a date, DeepInfra forwards requests to a recommended replacement after at least one week of notice, a silent substitution that keeps your requests running past the retirement date.
Across Copilot and Cursor, “handles churn well” means a documented, dated deprecation list plus a pinned-model path.
Where do you find each vendor’s model deprecation schedule and lifecycle policy, and is there a machine-readable feed?
Each vendor publishes its schedule in a different place. Only Bedrock, Microsoft and OpenAI offer a machine-readable feed. OpenAI’s deprecations page carries shutdown dates, and its List models API exposes shutdown_date. Anthropic’s deprecations page carries retirement dates and the 60-day commitment. Google splits across the Gemini API deprecations page and a separate Vertex AI calendar. Bedrock has a Model lifecycle page plus per-model cards, and Foundry has a retirement schedule and a Models API.
The machine-readable options are Bedrock’s modelLifecycle field, Microsoft’s lifecycleStatus and deprecation.inference fields, and OpenAI’s shutdown_date, though Microsoft’s Deprecated means retired, while Deprecating means still serving. Wiring one of these feeds into CI catches a retirement before it breaks production. Pair it with a quarterly retirement calendar of pinned model IDs, since no vendor-neutral calendar exists. The calendar is only useful if you pin dated IDs.
The stated notice and the delivered runway are two different numbers, and the gap is the figure you have to plan around. The runway is also channel-specific, because Bedrock and Foundry each publish their own retirement dates that differ from the first-party schedule. Your runway is the notice window minus your re-certification time. The practical response is a machine-readable feed in CI, a quarterly retirement calendar, and a contract that makes the runway binding, which is the negotiation that gets it in writing.
Frequently Asked Questions
Is “deprecated” the same as “retired”?
No, and the labels are not consistent across vendors. Anthropic runs four lifecycle states, Active, Legacy, Deprecated and Retired, where a deprecated model still serves but a retired one stops. OpenAI and Google use different terminology again, and Google only publishes “earliest possible” shutdown dates. Read each vendor’s lifecycle definitions before you treat any label as a deadline.
What actually breaks when a model is retired?
The endpoint stops serving requests. On Microsoft Foundry a retired model returns a 410 Gone error, while on OpenAI self-serve fine-tuning closes days before the model itself, so two things can fail on two dates. Separately, a model can change while still alive, as Claude Opus 5.5 did with token counts, response shape and sampling parameters.
Do fine-tuned models retire along with their base model?
Yes. Fine-tuned models follow the base model’s retirement schedule, and the fine-tuning service itself can close earlier. When OpenAI retired GPT-4o and GPT-4o mini in January 2027, self-serve fine-tuning ended on 6 January, a week ahead of the 14 January model retirement, so a fine-tuned deployment goes down in two stages.
Is it safe to point my app at the latest model alias?
No. The latest alias is a moving target, not a pin. Google gives only about two weeks’ notice for changes to latest and preview models, so your code can be swapped to a different model with almost no runway. Pin dated model IDs in production and treat the alias as a convenience for experiments.
Does buying through a cloud marketplace always give me more time?
Not always. Amazon Bedrock’s Legacy period can extend runway past the first-party retirement date, but Microsoft Foundry dates “aren’t extendable” and it auto-upgrades standard deployments unless you set NoAutoUpgrade. A marketplace layers a second clock and a second set of terms, so it is a buffer on some platforms and an extra failure point on others.
What exactly is a Legacy period or extended access window?
It is the transitioned or paid window after the public retirement date, during which a model still serves at provider-set pricing. Amazon Bedrock offers a Legacy period, with a 15-day inactivity clause that can cut it short. Check whether the window exists, what it costs, and whether it is guaranteed before you rely on it.
What is a change-of-control clause, and why should I care?
It is a contract term that lets a supplier wind down a service after an acquisition. When SpaceX acquired xAI, Cursor’s OpenAI model supply was proposed to end under a change-of-control clause, leaving users a roughly 76-day “proposed” window. The clause turns an ownership event into your deprecation risk, so it belongs in your own supplier contracts.
Does bring-your-own-key (BYOK) protect me from deprecation churn?
Only partly. On Cursor, BYOK covers local Chat and Agent but not Tab, Auto, agents, Automations, CLI or SDK, so most of the product still runs on the supplier’s models. Zero Data Retention also does not travel with a BYOK key. BYOK shifts some control to you but leaves the churn decision with the supplier.
Are open-weight models immune to deprecation?
No. Self-hosting an open-weight model relocates the exit cost rather than removing it. You own the serving stack and the update schedule, but you also inherit maintenance, security patching and your own retirement decisions. The vendor’s notice period disappears, and a different set of obligations takes its place.
What should I ask for in my contract to protect my runway?
Turn the notice period into binding language. Ask for a no-earlier-than retirement date, a minimum written notice clause, a defined change-of-control term, and a priced extended-access window if one exists. A commitment only pays off once it becomes a contractual obligation, not a policy page.
What happens if I miss a deprecation notice?
Your production breaks on a date you did not see coming, and you migrate with zero or negative runway. A six-month notice is worthless if your re-certification cycle takes seven months. Wire the lifecycle feed into CI and check the retirement calendar quarterly so the first warning is not a 410 Gone error.