The one-line answer: if you build on someone else's model, the time between a retirement announcement and your app throwing errors is set by the provider — and the providers disagree by more than an order of magnitude. Anthropic commits to at least 60 days. OpenAI documents at least 6 months for its GA models. Google publishes no guaranteed notice period for stable Gemini models at all — only ~2 weeks for previews. That gap is your migration budget, and you don't get to choose it. Here's what each provider actually promises, and the one number that should drive your planning.
The table, first (because this is the part you came for)#
| Provider | Documented notice floor | Versioning | Where it's published |
|---|---|---|---|
| Anthropic (Claude API) | ≥ 60 days for public models; active models "not sooner than" ~12 mo | Alias + dated snapshot — pin the date | platform.claude.com |
| OpenAI | ≥ 6 mo (GA); ≥ 3 mo (GA variants); ~2 wk (preview) | Snapshot pins; alias / -latest auto-updates | developers.openai.com |
| Google (Gemini/Vertex) | No quantified floor for stable; ~2 wk for preview & -latest | Stable 2.5+ has no auto-alias — pin the name | ai.google.dev |
| Azure OpenAI | ≥ 60 days (GA); 12-mo min availability | Dated versions; "not sooner than" date | learn.microsoft.com |
| AWS Bedrock | ≥ 6 mo after entering "Legacy" | Active → Legacy → EOL | docs.aws.amazon.com (Bedrock lifecycle) |
| DeepSeek | No published notice policy | Legacy aliases → explicit V4 IDs | deepseek.ai / API docs |
A note on trust, because this is a facts piece: Anthropic's numbers below are confirmed against its live deprecations page (August 6, 2026). The OpenAI, Google, and Azure figures reflect each provider's published policy language as captured this week; the exact wording moves, so re-read the linked page before you hard-code a plan around a specific number. That caveat is itself the lesson — the deprecations page is the source of truth, and you should be watching it.
Anthropic: 60 days, in writing — and a 12-month runway in practice#
Anthropic's deprecations page states it plainly: the company notifies customers with active deployments and provides "at least 60 days' notice before model retirement for publicly released models." In practice the runway is usually much longer — active models are listed with retirement dates phrased as "not sooner than" roughly 12 months out, so you typically see the end coming a year ahead.
The recent history bears out the cadence. claude-opus-4-1-20250805 was deprecated on June 5, 2026 and retired on August 5, 2026 — two months of notice, exactly the floor. claude-sonnet-4 and claude-opus-4 were retired June 15, 2026; the 3.x family wound down through the first half of the year. Anthropic uses a four-state lifecycle — Active → Legacy → Deprecated → Retired — and once a model is Retired, "requests to retired models will fail." (Worth knowing, if you pin sampling params: temperature, top_p, and top_k are already deprecated on Opus 4.7 and later, returning a 400 when set to a non-default value.)
One scope trap: Anthropic's dates apply to Anthropic-operated platforms — the Claude API, Claude on AWS, and Microsoft Foundry. It explicitly notes that partner-operated platforms (Amazon Bedrock, Google Cloud) set their own schedules, so the same model can have different lifecycle dates depending on where you call it.
OpenAI: the longest documented floor — if you're on a GA model#
OpenAI's deprecations page documents the most generous guaranteed floor of the majors, but it's tiered:
- Generally available models: at least 6 months before shutdown.
- GA variants — the
-chat-latestchat models,-codexcoding variants,-deep-researchmodels — at least 3 months. - Preview models (anything with "preview" in the name): as little as ~2 weeks, and OpenAI reserves the right to move faster for safety or compliance reasons, "with as much notice as reasonably possible."
The versioning model is the familiar dual system: a dated snapshot like gpt-4o-2024-08-06 is frozen and won't change under you, while the bare alias (gpt-4o) and any -latest name auto-update to the newest default. That auto-update is a silent model swap — the danger we walk through in the companion runbook below. OpenAI runs periodic shutdown waves of older snapshots and notifies impacted customers by email in addition to the deprecations page.
Google: the outlier — no promised runway for stable models#
Here's the finding that should change how you plan. For stable Gemini models, Google does not publish a guaranteed minimum notice period. What it does commit to is narrow:
- Preview models: ~2 weeks' notice before deprecation.
-latestalias breaking changes: ~2 weeks' notice by email before the version behind-latestchanges.
For stable models, Google's docs describe the listed shutdown dates as the "earliest possible dates on which a model might be retired," with the exact date communicated "with advance notice" — but no committed number of months. Compared with OpenAI's 6 months and Anthropic's 60 days, that's a genuine, documented difference, not a technicality.
Google's versioning cuts the other way, though, and it helps you: Gemini 2.5-and-later stable models dropped the auto-updating alias entirely. There's no floating name to get hot-swapped — you pin the explicit stable model name, and Google recommends exactly that for production. Reserve -latest (which can resolve to stable, preview, or experimental) for when you've chosen "always newest" deliberately. Two more Google-specific traps: retirement dates differ between the Gemini API / AI Studio and Vertex AI for the same model, and the concrete example is instructive — Gemini 2.0 Flash and Flash-Lite shut down on June 1, 2026, and gemini-flash-latest, which had resolved to 2.0 Flash, now returns a 404.
Azure OpenAI and AWS Bedrock: the marketplace layer has its own clock#
If you reach OpenAI's models through Azure, the policy is Microsoft's, not OpenAI's: at least 60 days' notice before a GA model retirement, ~30 days before a preview version upgrade, and a 12-month minimum availability window for a GA model version, with the "not sooner than" retirement date set at launch. AWS Bedrock uses an Active → Legacy → EOL lifecycle and commits to at least 6 months of access once a model enters Legacy (with an extended-access phase that may carry higher, provider-set pricing). The takeaway from the Anthropic scope note applies here in reverse: check the marketplace's own deprecations page, because it is the authority for models you consume there.
DeepSeek and the open-weight vendors: assume no notice#
DeepSeek publishes no formal notice-period commitment we could find, and its recent behavior sets expectations: the legacy deepseek-chat and deepseek-reasoner aliases — backward-compat shims that had resolved to deepseek-v4-flash — were hard-retired at 15:59 UTC on July 24, 2026, with no grace period and no soft redirect. After the cutoff, only the explicit deepseek-v4-flash and deepseek-v4-pro IDs resolve; the old names return errors. For the fast-moving open-weight labs, treat every alias as retirable on short notice and pin explicit versions.
What to actually do with this#
The planning rule falls straight out of the table: size your migration runway to the shortest promise among the providers you actually depend on. If your stack is all Anthropic GA models, you've got a comfortable ~12-month practical runway with a 60-day hard floor. If it includes an OpenAI preview model or any Google stable model, your real guaranteed notice drops toward two weeks — or, for Google stable, toward "whenever they decide." That shortest number is the one your on-call rotation lives with.
Concretely: pin dated snapshots (or explicit stable names on Google), keep model IDs in one place, subscribe to every provider's deprecations page, and hold a golden eval set so you can qualify a replacement fast. We put the full engineering runbook — the grep to find every model ID, the config pattern, the fallback chain for the hard-retirement case — in How to Survive a Model Retirement. For the deeper look at picking between specific frontier models when you do have to move, see Claude Opus 5 vs Kimi K3 for agentic coding. The policies will keep changing; the discipline of watching them shouldn't.
