Twelve Claude models are retired and their identifiers no longer respond on the Claude API. The four retired in 2026 are Claude Opus 4.1, on 5 August 2026; Claude Opus 4 and Claude Sonnet 4, on 15 June 2026; and Claude Haiku 3, on 20 April 2026. Anthropic publishes an official replacement for each one.
Data last verified: 30 August 2026, against the official Anthropic documentation linked at the end.
If you are reading this, something that worked has probably stopped working. It usually shows up in an ugly way: the automation does not raise an alarm, it just stops producing, and the error sits in a log nobody reads. This is about identifying which model died, which one to move to, and what else breaks besides the name.
Which Claude models are retired, and when did each stop working?
These were retired during 2026. They are the ones breaking integrations that were live.
| API identifier | Retirement date | Official replacement |
|---|---|---|
claude-opus-4-1-20250805 | 5 August 2026 | Claude Opus 4.8 |
claude-opus-4-20250514 | 15 June 2026 | Claude Opus 4.8 |
claude-sonnet-4-20250514 | 15 June 2026 | Claude Sonnet 4.6 |
claude-3-haiku-20240307 | 20 April 2026 | Claude Haiku 4.5 |
claude-3-7-sonnet-20250219 | 19 February 2026 | Claude Sonnet 4.6 |
claude-3-5-haiku-20241022 | 19 February 2026 | Claude Haiku 4.5 |
claude-3-opus-20240229 | 5 January 2026 | Claude Opus 4.8 |
These have been retired for longer. Listed in case you maintain old code — but if your integration was still calling one of them, it has been broken for far longer than you think.
| API identifier | Retirement date | Official replacement |
|---|---|---|
claude-3-5-sonnet-20241022 | 28 October 2025 | Claude Sonnet 4.6 |
claude-3-5-sonnet-20240620 | 28 October 2025 | Claude Sonnet 4.6 |
claude-3-sonnet-20240229 | 21 July 2025 | Claude Sonnet 4.6 |
claude-2.1 | 21 July 2025 | Claude Opus 4.8 |
claude-2.0 | 21 July 2025 | Claude Opus 4.8 |
Why does it still work for my colleague and not for me?
Because not every platform retires at the same time, and this trips people up. The dates above apply to the platforms Anthropic operates: the Claude API, Claude Platform on AWS and Microsoft Foundry. Amazon Bedrock and Google Cloud set their own schedules, so a model can be dead in one place and still serving in another.
Which is why it is worth not writing "the model no longer exists" in an internal report. The accurate version is "that identifier no longer responds on the platform we use". If your provider says it still works, they are probably not lying to you — you are looking at different platforms.
What else breaks besides the model identifier?
This is the part that eats the hours, because changing the model name is the easy bit. From Claude Opus 4.7 onwards, temperature, top_p, top_k are deprecated: they return a 400 error when set to a non-default value. The official recommendation is to omit them and steer behaviour through the prompt.
So a migration that looked like find-and-replace turns into reviewing every call. And one detail catches people out: the Python SDK 1.0 and later removes them from the signature, so passing them raises a TypeError instead of reaching the API — meaning the failure surfaces in different places depending on how you call.
Is migrating more expensive?
It depends on the pair, but when it drops, it drops a lot. Using an ordinary small-business scenario — an agent handling 500 conversations a month, about 3,700 tokens per conversation — this is what changes in the model bill:
| Migration | Before | After | Difference |
|---|---|---|---|
| Claude Opus 4.1 → Claude Opus 4.8 | $48.75 per month | $16.25 per month | down $32.50 (€27.78) |
| Claude Opus 4 → Claude Opus 4.8 | $48.75 per month | $16.25 per month | down $32.50 (€27.78) |
| Claude Sonnet 4 → Claude Sonnet 4.6 | $9.75 per month | $9.75 per month | same price |
| Claude Haiku 3.5 → Claude Haiku 4.5 | $2.60 per month | $3.25 per month | up $0.65 (€0.56) |
The striking case is Claude Opus 4.1: moving to Claude Opus 4.8 leaves that agent at $16.25 a month instead of $48.75, a third of what it cost. If you have had the migration parked for months, you have been paying for it.
And the honest case is Claude Haiku 3.5, where the replacement costs slightly more. It is cents, but it gets said: not every migration saves money.
European Central Bank reference rate for 21 August 2026: €1 = $1.1699. Dollars are what Anthropic bills. For the current price breakdown, see the article on Claude Sonnet 5 pricing.
Before you migrate, check the small print on the new model
A model change can bring different data-handling conditions, and that matters in a law firm or a clinic. The concrete example: Claude Fable 5 requires 30-day data retention and is not available under zero data retention. If your organisation has ZDR, requests to that model return 400 invalid_request_error.
There is a way out that does not mean giving up ZDR across the organisation: 30-day retention can be enabled for a single workspace, leaving the rest as it is. But that is a decision to take knowingly, not to discover when the API starts returning errors.
What we do about this
At AizuaLabs we maintain AI agents for small and mid-sized companies, and part of the job is exactly this: making sure the client never notices the day a model is retired. We pick the model case by case, keep the identifier in one place rather than scattered through the code, and watch the retirement calendar instead of waiting for the error.
If you have an automation stopped and no idea where to start, the quick route is a free 60-minute audit: we look at what broke and tell you what it costs to fix, whether or not you work with us. If you would rather see how we build things first, here is how we build AI agents for companies.
Frequently asked questions
What error does a retired Claude model return?
The request fails: the identifier is no longer served on that platform. The official documentation puts it as requests to retired models will fail. Check the exact identifier against the deprecations page before assuming the problem is something else.
Which model should I migrate to from Claude Opus 4.1?
Claude Opus 4.8, the official replacement. It is also cheaper: for an agent handling 500 conversations a month, model cost goes from $48.75 to $16.25 per month.
When was Claude Opus 4.1 retired?
On 5 August 2026 on Anthropic-operated platforms. It was announced on 5 June 2026. Amazon Bedrock and Google Cloud set their own dates.
Why does my integration still work with a retired model?
Most likely because you call through Amazon Bedrock or Google Cloud, which have their own retirement schedules. Anthropic's published dates apply to the Claude API, Claude Platform on AWS and Microsoft Foundry.
Is changing the model name enough to migrate?
No. From Claude Opus 4.7 onwards, temperature, top_p, top_k are deprecated and return a 400 error when set to a non-default value. The official recommendation is to omit them and steer behaviour through the prompt.
Is migrating expensive?
The price change depends on the pair. Claude Opus 4.1 to Claude Opus 4.8 cuts model cost to a third; Claude Haiku 3.5 to Claude Haiku 4.5 goes up slightly. What usually costs is not the model but reviewing the calls and testing again.
How do I stop this happening again?
Keep the model identifier in one place in your code, check the official deprecations page on a schedule, and make sure integration errors reach somewhere a human reads them. Most of these failures are not spotted through an error but because someone notices results are missing.
Sources
- Anthropic — Model deprecations. Lifecycle status, retirement dates, official replacements and deprecated parameters. Accessed 30 August 2026.
- Anthropic — Pricing. Pricing for the retired models still listed and for their replacements. Accessed 30 August 2026.
- Anthropic — API and data retention. 30-day retention and zero-data-retention availability. Accessed 30 August 2026.
- European Central Bank — Euro reference exchange rates. €1 = $1.1699 on 21 August 2026.
We revise this article whenever Anthropic announces a new retirement. If a date looks wrong, tell us and we will correct it.