What you will be able to do
- Name the four model lifecycle states Anthropic uses and say what each one means for a running application
- Plan a migration after a deprecation notice using the notice period, the Console usage audit and early testing
- Explain why a model on Amazon Bedrock or Google Cloud can have a different retirement schedule from the same model on the Claude API
- Describe what Anthropic promises to keep stable within a Messages API version, and what it may still add or change
Key concept
Model lifecycle — Every Claude model moves through the states Active, Legacy, Deprecated and Retired. Maintaining a Claude application means tracking which state your pinned model is in and moving to a replacement before it reaches Retired, because at that point requests stop working.
1.The four states a Claude model moves through
The systems life cycle of a Claude-powered application does not end when it ships. Operating and maintaining it means living with a dependency that changes under you: the model. Anthropic retires older models regularly as newer ones launch, and says openly that applications relying on its models may need occasional updates to keep working. To plan those updates, you need the vocabulary Anthropic uses for where a model is in its life.
| State | What it means | Example from the deprecations table |
|---|---|---|
| Active | Fully supported and recommended for use | claude-opus-5-5 (not sooner than September 22, 2027) |
| Legacy | No longer receives updates; may be deprecated in the future | No model is listed in this state in the current table |
| Deprecated | Still works but is no longer recommended; a replacement and a retirement date are assigned | claude-mythos-preview (deprecated June 9, 2026) |
| Retired | No longer available; requests fail | claude-3-haiku-20240307 (retired April 20, 2026) |
Notice what each transition means for you. Moving to Legacy only means the model stops receiving updates. Moving to Deprecated changes nothing in how your calls behave: the model keeps answering, but a retirement date now exists and a recommended replacement is named. Only Retired breaks things. That is why the retirement date, not the deprecation date, is the deadline in your plan. The "tentative retirement date" column for Active models works as a floor. It says "not sooner than" a given date, so treat it as the earliest point to plan around, not a promise of when retirement will happen.
Sources1
2.Responding to a deprecation notice
You are not left to discover deprecations yourself. Anthropic emails impacted customers and records every deprecation in the documentation. For publicly released models it guarantees a minimum notice period before retirement. Anthropic's own recommended steps turn that notice period into a plan: check the documentation regularly for new deprecations, test your applications with newer models well before your current model's retirement date, update your code to the recommended replacement as soon as possible, and contact support if you need help migrating.
The hard part in a real organisation is often finding every place the old model is still called. Several services, scripts and API keys can each hardcode a model name. The Claude Console includes an audit for this: on the Usage page, click Export, and the downloaded CSV breaks usage down by API key and model. That list shows exactly which keys still send traffic to the deprecated model, so you can prioritise the migration before the retirement date instead of waiting for failures to tell you.
A platform engineering team runs a customer support application on Claude in production and wants to build a compliance policy for planning migration testing whenever a model they depend on is deprecated. Per Anthropic's stated commitment, what is the minimum notice period they can count on before a publicly released model they are actively using is retired?
Correct answer: B — At least 60 days before the retirement date
- A. Incorrect -- Anthropic's stated minimum commitment is 60 days, not 30, so a policy built around 30 days would assume less guaranteed lead time than Anthropic actually provides.
- B. Correct -- Anthropic notifies customers with active deployments at least 60 days before a publicly released model's retirement date.
- C. Incorrect -- 90 days overstates the guaranteed minimum; actual notice in specific cases has sometimes exceeded 60 days, but the committed floor is 60 days, not 90.
- D. Incorrect -- 180 days is well beyond the guaranteed minimum notice window Anthropic commits to for publicly released models.
Sources1
3.Partner platforms keep their own calendars
The dates on the deprecations page cover Anthropic-operated platforms: the Claude API, Claude Platform on AWS, and Microsoft Foundry. The same model served through Amazon Bedrock or Google Cloud is on a different calendar. Those partners run their own retirement schedules, so a model's status and dates there can differ from what Anthropic lists for its own platforms.
This matters for any team that runs the same application on more than one platform, for example the Claude API for most customers and Bedrock for customers who need a particular region. One deprecation announcement does not give you one deadline. Before assuming your Bedrock or Google Cloud deployment retires on the same date, check that partner's own model table. Anthropic's page links to separate Amazon Bedrock and Google Cloud tables for exactly this reason. Your migration plan may need a different date for each platform.
Ahead of an upcoming retirement date, a platform team needs to find every place in their production traffic where a soon-to-be-retired model is still being called, broken down by API key and model, before they finalize their migration plan. Which action should they take in Claude Console?
Correct answer: C — Open the Usage page in Claude Console and export a CSV broken down by API key and model.
- A. Incorrect -- this bypasses the self-service audit tool Anthropic already provides and adds unnecessary delay compared to exporting usage data directly.
- B. Incorrect -- extended thinking is a reasoning feature unrelated to usage auditing and does not add model identifiers to a response payload for this purpose.
- C. Correct -- the Usage page's CSV export is the documented way to audit API usage broken down by API key and model, making it straightforward to locate deprecated-model traffic.
- D. Incorrect -- searching internal logs for a literal string depends on logging conventions the team controls and won't reliably capture every call broken down by API key and model the way the Console export does.
Sources1
4.API versions: what stays stable and what can grow
Models are one moving part. The API contract you call them through is another, and it follows its own versioning scheme. Within a given version of the Messages API, Anthropic promises to keep your existing request and response fields intact, while reserving the right to add things around them.
| Anthropic preserves | Anthropic may |
|---|---|
| Existing input parameters | Add additional optional inputs |
| Existing output parameters | Add additional values to the output |
| — | Change conditions for specific error types |
| — | Add new variants to enum-like output values (for example, streaming event types) |
No. New variants of enum-like output values, streaming event types included, are explicitly allowed within the same API version. A client that crashes on an unknown event type can break without Anthropic breaking its contract. Handle unknown values gracefully, for example by ignoring them.
The version history shows what a real version bump looks like. The 2023-06-01 version changed streaming server-sent events so that completions arrive incrementally as named events, and it removed the old data: [DONE] event. Anthropic recommends using the latest API version whenever possible. Older versions are treated as deprecated and may not be available to new users.
Sources2
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Once a model is marked Deprecated, requests to it start failing, so an outage is already under way.Why is that wrong?
A deprecated model keeps working. It gets a recommended replacement and a retirement date, and requests fail only once the model is Retired.
Covered in The four states a Claude model moves through
2.A retirement date announced for the Claude API automatically applies to the same model on Amazon Bedrock.Why is that wrong?
Anthropic's dates cover Anthropic-operated platforms only. Amazon Bedrock and Google Cloud set their own schedules, so check each partner's own table.
Covered in Partner platforms keep their own calendars
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Legacy: The model will no longer receive updates and may be deprecated in the future.”
↩︎ The four states a Claude model moves through“Retired: The model is no longer available for use. Requests to retired models will fail.”
↩︎ The four states a Claude model moves through“providing at least 60 days' notice before model retirement for publicly released models.”
↩︎ Responding to a deprecation notice“Test your applications with newer models well before the retirement date of your current model.”
↩︎ Responding to a deprecation notice“Review the downloaded CSV to see usage broken down by API key and model.”
↩︎ Responding to a deprecation notice“Partner-operated platforms (Amazon Bedrock and Google Cloud) set their own retirement schedules, so a model's lifecycle status and dates can differ.”
↩︎ Partner platforms keep their own calendars“Retired: The model is no longer available for use. Requests to retired models will fail.”
↩︎ Key concept“Deprecated: The model is still functional but no longer recommended.”
↩︎ Exam trap 1“Partner-operated platforms (Amazon Bedrock and Google Cloud) set their own retirement schedules, so a model's lifecycle status and dates can differ.”
↩︎ Exam trap 2 - 2.https://platform.claude.com/docs/en/api/versioningOfficial docs
“Generally, if you are using the API as documented in this reference, Anthropic will not break your usage.”
↩︎ API versions: what stays stable and what can grow“Add new variants to enum-like output values (for example, streaming event types)”
↩︎ API versions: what stays stable and what can grow“Anthropic recommends using the latest API version whenever possible.”
↩︎ API versions: what stays stable and what can grow