Sonnet 4.5 is retiring. Here's what that actually means
Anthropic retires Sonnet 4.5 on November 30. Bedrock says April 8, Vertex has not said at all, and none of the dates is the one in the announcement.
In August I started building Token Stanza so I could keep talking to Claude Sonnet 4.5. On September 30, Anthropic announced its retirement. When I went to check the dates, I found that “retiring” means three different things on the three platforms that serve the model, and the dates are not the same. This post is what I found when I read each one, and what it means if you depend on the model.
| Platform | Status change | Price increase | Last day |
|---|---|---|---|
| Claude API | Deprecated Sept 30 | — | Nov 30, 2026 |
| Amazon Bedrock | Legacy Oct 8 | Jan 8, 2027 | April 8, 2027 |
| Vertex AI | None announced | — | Unknown |
Here’s what those dates mean in practice. On the Claude API and Bedrock, requests to the model fail after the retirement or end-of-life date. Nothing about a conversation’s stored history changes—what changes is which model can be asked to continue it. The rest of this post is what each platform means by those words.
What Anthropic announced
On September 30, 2026, Anthropic deprecated claude-sonnet-4-5-20250929 on the Claude API. The retirement date is November 30, 2026, and claude-sonnet-5-5 is the recommended replacement (Anthropic’s model deprecations page). Anthropic’s own words for the lifecycle are Active, Legacy, Deprecated and Retired. Deprecated means the model still works but is not recommended and has a retirement date. Retired means requests fail. The policy promises at least 60 days’ notice for publicly released models. The page says plainly that its dates apply to Anthropic-operated platforms only.
What happens to the weights
One line on that page is worth knowing if you are the kind of person who wants to keep a model. Anthropic has committed to preserving the weights of every publicly released model for at least the company’s lifetime, and says it hopes to make past models available again some day. That is not access, but it is not deletion either.
What stops and what doesn’t
A retired model fails requests but leaves history alone. In Token Stanza that line is concrete. A conversation keeps its history, its title and its model name after the model is gone. It can still be read, and it exports as Markdown or JSON on its own, or inside the account zip from Settings with every other conversation and its handoff summary. What it cannot do is take a new message, because that message would go to a model the platform no longer serves. The last section of this post is about where a new message goes instead.
Bedrock keeps its own calendar
Bedrock’s lifecycle page for models launched before September 7, 2026 (Model lifecycle (Legacy)) lists Claude Sonnet 4.5 as Legacy from October 8, 2026. Public extended access starts on January 8, 2027, and end of life is April 8, 2027, across every region it names. Legacy means new customers cannot adopt the model and existing customers may lose access after 15 days of inactivity. Extended access means higher pricing, set by the provider; AWS hasn’t announced the rate. End of life means requests fail. The page says in as many words that its dates are Bedrock’s and may differ from Anthropic’s. There is precedent. Claude Sonnet 4 left the Claude API on June 15, 2026 and Bedrock on October 14, four months later.
The region-by-region picture
Bedrock’s dates are per region, and the page says a model can stay Active in regions its row does not list. The app takes a daily snapshot of every Anthropic model in every region it can reach, with the lifecycle dates, and records what changed since the day before. That snapshot design—watching for daily diffs, looking for regions diverging—catches the specific problem with cross-region inference profiles.
For Sonnet 4.5 the table is boring today: every region reports the same three dates. That is the expected shape, and it is the reason the snapshot exists. The moment one region’s date moves ahead of the others, a conversation served through a cross-region inference profile could land in a region that has already stopped serving the model. The caller does not pick the region. If one region a profile spans drops the model while the others keep it, some requests could fail and some succeed, with nothing on the caller’s side to say which. The daily diff is what would tell me. A later post in this series is about watching the catalog rather than reading it once.
The snapshot also shows something the lifecycle page does not. In every region, Sonnet 4.5 is reachable only through an inference profile, never by a direct model id. A request for Sonnet 4.5 is never a request to one region. It is a request to Bedrock to pick one. That is the point about per-region dates above, and it is also why the prompt-caching post has a paragraph about cache misses that are nobody’s fault.
Vertex has not said
Google is the odd one out: it has not announced a date at all. Its partner-model deprecations page lists six Claude models with a deprecation date and a shutdown date each, from Claude 3 Opus through Claude 3 Haiku. Sonnet 4.5 is not among them. The model’s own page says “Retirement date not sooner than: September 29, 2026”. That is a floor, not a date, and the floor has already passed. Read literally, Google has reserved the right to retire the model at any time and has not yet exercised it.
There is a pattern in the six entries that are there. It is worth knowing even though it is not a promise. Five of the six put six months between the deprecation date and the shutdown date. Claude 3.7 Sonnet was deprecated on November 11, 2025 and shut down on May 11, 2026. Claude 3.5 Haiku: January 5 and July 5, 2026. Claude 3 Haiku: February 23 and August 23, 2026. The exception is the oldest, Claude 3 Opus, which got five weeks. If Google follows its usual pattern, a Sonnet 4.5 deprecation announced today would mean a shutdown next spring, which is where Bedrock already is. If you are on Vertex, the thing to watch is the deprecations page. The model page will not change until it does.
Why I still call it Vertex
Google has renamed the product. The pages that were Vertex AI are now “Gemini Enterprise Agent Platform” (the release note is dated April 22, 2026), and the old URLs redirect. The model ids and the endpoints are the same, and Anthropic’s own docs call it Claude Models on Google Cloud (Google managed). I mention this because you’ll see both names in documentation; they mean the same platform. I keep calling it Vertex, because that is what the app’s provider is named, and that name is stored on every conversation.
What to do if you depend on it
Four things, in the order I would do them.
- Put the date that applies to you in a calendar with a few weeks’ margin. November 30 on the Claude API. On Bedrock, January 8 for the price change and April 8 for the end. On Vertex there is no date, so the entry is a weekly look at the deprecations page.
- Test the replacement before the date, not after, on the requests you actually send.
- Audit which requests still name the old id. On Bedrock, know which regions your inference profile spans, since the dates are per region.
- Export anything you want a copy of while the model is still being served. In Token Stanza the history stays either way, and the export is there if you want one.
If you are on Bedrock you have until April, not November. If you are on Vertex you have no date yet. That is worse, not better, because the notice could come any day.
What Token Stanza will do
For Token Stanza, a retirement is a routing decision. A conversation is pinned to one provider and one model for its life, and the pinning is what makes the prompt cache work. So the decision is which model new conversations start on, and what happens to conversations still on the old one.
The plan is to keep Sonnet 4.5 for as long as any platform serves it. The app talks to three. That was never about redundancy. It was about exactly this, from the beginning. After November 30, new Sonnet 4.5 conversations will keep starting on Bedrock, which has committed to April 8 and will charge extended-access pricing from January 8. Or they will start on Vertex, if Google’s silence turns out to mean a later date. A conversation already running on a platform that retires the model cannot continue there, because the provider is part of what it is pinned to. Its history stays readable and exportable. The handoff summary is the way to carry its thread into a new conversation on a platform that still has the model.
When the last platform stops, new conversations move to Sonnet 4.6 rather than jumping a generation to 5.5. It is the nearest model to the one the app was built around. It is on the same calendar. Anthropic’s deprecations page lists Sonnet 4.6 with a tentative retirement of not sooner than February 17, 2027, which is before Bedrock stops serving Sonnet 4.5 on April 8. The successor could leave the Claude API before the predecessor leaves Bedrock. That move is not free on my side. Sonnet 4.6 rejects the budget_tokens thinking configuration that every 4.5 turn sends, and wants adaptive thinking with an effort level instead. That means the thinking request every turn sends, and the controls behind the composer that set it, have to be rewired for that model before the first conversation starts on it. The app already has that path built for the model chat on the Anthropic API, where the newer models already run. After November 30 this post gets a dated note at the top with what actually happened, and another if Google names a date. The Updated date beside the title will say when.