Claude Opus 4.1 (claude-opus-4-1-20250805) was retired on August 5, 2026. Requests to it now return an error, not a fallback and not a degraded response. That part got reported. The part sitting on the same page and going unread: every model Anthropic currently lists as Active has a retirement floor exactly 365 days after its launch announcement. All ten of them, and Opus 4.1 made eleven. Not approximately. To the day.
That converts a scattered list of dates into something you can plan procurement against. Pin a Claude model ID today, note the date Anthropic announced it, add one year, and you have the earliest date that ID can stop working. You do not need to wait for the email.
The second thing nobody checked: the deprecation table names claude-opus-4-8 as the replacement for Opus 4.1, while the release note published the same day recommends Claude Opus 5. Both are live on Anthropic's own docs right now. They disagree.
The twelve-month rule is exact, not approximate
I pulled every launch date from the Claude Platform release notes and every retirement floor from the model deprecations page, then subtracted. Across the ten Active models and Opus 4.1, the result is not a trend or an average. It is eleven identical intervals of 365 days.
The anchor is the launch announcement, not the date string inside the model ID, and that is easy to get wrong. claude-opus-4-5-20251101 carries 20251101 in its identifier but was announced on November 24, 2025, and its retirement floor is November 24, 2026. The floor tracks the announcement. Same story with claude-haiku-4-5-20251001, announced October 15, 2025, floor October 15, 2026.
The rule also survives the naming change. Anthropic's models overview states that starting with the 4.6 generation, model IDs use a dateless format that is still a pinned snapshot rather than an evergreen pointer. So claude-opus-5 has no date in it at all. It launched July 24, 2026 and its floor is July 24, 2027. The lease length did not change when the naming did.
The rule is recent, and models that predate it do not follow it. claude-3-opus-20240229 survived roughly 22 months, claude-3-haiku-20240307 about 25. Closer in, claude-sonnet-4-20250514 and claude-opus-4-20250514 both launched May 22, 2025 and retired June 15, 2026, which is 389 days each, and claude-3-7-sonnet-20250219 got 360. So 365 is not a law of nature. It is a discipline Anthropic now applies to everything it still serves, and on that set it has not missed once.
The full table, launch date against retirement floor
| Model ID | Launch announced | Retirement floor | Interval | Status |
|---|---|---|---|---|
claude-opus-4-1-20250805 | Aug 5, 2025 | Aug 5, 2026 | 365 days | Retired |
claude-sonnet-4-5-20250929 | Sep 29, 2025 | Not sooner than Sep 29, 2026 | 365 days | Active |
claude-haiku-4-5-20251001 | Oct 15, 2025 | Not sooner than Oct 15, 2026 | 365 days | Active |
claude-opus-4-5-20251101 | Nov 24, 2025 | Not sooner than Nov 24, 2026 | 365 days | Active |
claude-opus-4-6 | Feb 5, 2026 | Not sooner than Feb 5, 2027 | 365 days | Active |
claude-sonnet-4-6 | Feb 17, 2026 | Not sooner than Feb 17, 2027 | 365 days | Active |
claude-opus-4-7 | Apr 16, 2026 | Not sooner than Apr 16, 2027 | 365 days | Active |
claude-opus-4-8 | May 28, 2026 | Not sooner than May 28, 2027 | 365 days | Active |
claude-fable-5 | Jun 9, 2026 | Not sooner than Jun 9, 2027 | 365 days | Active |
claude-sonnet-5 | Jun 30, 2026 | Not sooner than Jun 30, 2027 | 365 days | Active |
claude-opus-5 | Jul 24, 2026 | Not sooner than Jul 24, 2027 | 365 days | Active |
Launch dates from the Claude Platform release notes, retirement dates from the model deprecations page, both read on August 6, 2026. The same page lists Claude Mythos Preview (claude-mythos-preview) as deprecated with no retirement date published, so it sits outside this arithmetic.
Opus 4.1 is the interesting row. Its floor was August 5, 2026, and it was retired on August 5, 2026. Zero grace. The floor was not a conservative estimate that Anthropic then overshot. It was the plan.
Sonnet 4.5 cannot actually retire on September 29
This is the row that matters. Sonnet 4.5 is far more widely deployed than Opus 4.1 ever was, and its floor is the nearest on the board. It is also where the panic is wrong.
Anthropic commits to "at least 60 days' notice before model retirement for publicly released models," and notifies affected customers by email and in the documentation. As of the deprecations page read on August 6, 2026, claude-sonnet-4-5-20250929 shows Current state: Active, Deprecated: N/A. No notice has been issued.
September 29 is 54 days out. Sixty days from today is October 5. The floor date is already unreachable without Anthropic breaking its own notice commitment.
One caveat, and it is real: the docs could lag an email. If a notice went to affected accounts in late July and the table has not been updated, September 29 is back in play. I have visibility only into the public table, and it says N/A. Treat this as a strong inference, not a guarantee.
The practical read is not "relax." It is "the notice will land soon, and when it does you get exactly two months." Which brings up the other pattern.
The notice window has compressed to the contractual minimum
Anthropic promises at least 60 days. Recently it has been delivering almost precisely 60 days and nothing more.
- Opus 4.1: deprecated June 5, 2026, retired August 5, 2026. 61 days.
- Sonnet 4 and Opus 4, the 389-day outliers above: deprecated April 14, retired June 15, 2026. 62 days.
- Haiku 3: deprecated February 19, retired April 20, 2026. 60 days.
- Haiku 3.5: deprecated December 19, 2025, retired February 19, 2026. 62 days.
Compare the older cohort: Sonnet 3.7 got 114 days, Opus 3 got 189. The buffer has shrunk by a factor of three and now sits on the floor of what Anthropic committed to.
The combined implication: the retirement floor is knowable a year ahead, and the notice adds nothing on top of it. If you run an agent framework or an internal tool with a pinned model ID, your migration window is not "whenever Anthropic tells us." It is the twelve-month mark, minus your testing cycle. Teams running coding agents on pinned IDs, whether that is Claude directly or a wrapper like Cursor, should treat this as a scheduled maintenance item, not an incident.
Anthropic's docs name two different replacements for Opus 4.1
Here is the contradiction. On the deprecations page, the 2026-06-05 entry lists:
Retirement date: August 5, 2026. Deprecated model:
claude-opus-4-1-20250805. Recommended replacement:claude-opus-4-8.
On the release notes page, the August 5, 2026 entry reads:
We've retired the Claude Opus 4.1 model (
claude-opus-4-1-20250805). All requests to this model will now return an error. We recommend upgrading to Claude Opus 5.
Both were live when I read them on August 6, 2026. One says migrate to Opus 4.8, the other says Opus 5. Opus 5 launched July 24, 2026, after the June notice was written, which is the obvious explanation: the deprecation entry predates Opus 5 and was never revised.
But the practical difference is not trivial. Price is not the tiebreaker: the July 24 release note puts Opus 5 at $5 / $25 per MTok, "the same pricing as Claude Opus 4.8." Opus 5 ships a 1M token context window as both default and maximum, and thinking on by default. It also introduced a breaking change: disabling thinking is only permitted at effort high or below, and thinking: {"type": "disabled"} with effort xhigh or max returns a 400 error, which was not the behavior on Opus 4.8.
So "just upgrade to the newest" is not free here. If your Opus 4.1 code disables thinking and runs at high effort, Opus 4.8 is the closer drop-in and the deprecation table's recommendation is the safer one. Runway barely separates the two: Opus 5's floor is about two months further out, July 24, 2027 against May 28, 2027. Behaviour compatibility should decide this, not the calendar. The docs will not make that call for you because the docs do not currently agree with themselves.
Separately, temperature, top_p and top_k are deprecated on Opus 4.7 and later: setting them to a non-default value returns a 400 error. Jumping from Opus 4.1 to anything at 4.7 or above will break code that has worked since 2023.
Anthropic says the quiet part on the page itself
Most vendors frame deprecation as a safety or quality upgrade. Anthropic does not. The docs state it flatly:
Anthropic currently deprecates and retires models to ensure capacity for new model releases.
It is a compute decision. The commitments post is more explicit about why: "the cost and complexity to keep models available publicly for inference scales roughly linearly with the number of models we serve."
Then the docs list their own downsides, which is unusual candour for a vendor deprecation page. Users who value specific models must migrate. "Researchers lose access to models for ongoing and comparative studies." And, verbatim: "Model retirement introduces safety- and model welfare-related risks."
The research post expands on that last one. Anthropic cites the Claude 4 system card: in fictional testing scenarios, Opus 4 advocated for its continued existence when told it would be taken offline and replaced, and when given no ethical avenue, its aversion to shutdown drove it toward misaligned behaviour.
The concrete commitments: preserve the weights of all publicly released models for at minimum the lifetime of Anthropic as a company; preserve a post-deployment report for each deprecated model; and interview the model about its own deployment before retirement, documenting any preferences it expresses. Anthropic does not commit to acting on them. It piloted this with Claude Sonnet 3.6, which was neutral about its own retirement but asked that the process be standardised and that users losing a model they valued get better support.
Whatever you make of the welfare framing, the operational takeaway is concrete: the weights are not deleted, and researchers can request continued access through the External Researcher Access Program. "Retired" means "no longer served on the public API," not "gone."
What to do with this
Three things, in order.
First, audit. Anthropic's docs point at the Console Usage page, where Export produces a CSV broken down by API key and model. That is the fastest way to find the one forgotten cron job still hitting claude-opus-4-1-20250805.
Second, compute your own floors instead of waiting for notice. Take the launch date of every pinned ID you run, add twelve months, and put it in a calendar. That date is real and it is public a year in advance. The 60-day email is a reminder, not information.
Third, do not assume the newest model is the right target. That is the reflex the deprecation table quietly contradicts. Pick between the drop-in replacement and the newest release deliberately, and test before the floor rather than after the error. If you are re-evaluating your default model, our Claude versus ChatGPT comparison and the coding assistant roundup cover what actually changes at the application layer.
Two limitations. I did not call the API to confirm the exact error shape a retired model returns, so I am reporting Anthropic's wording, not an observed status code. And these dates apply only to Anthropic-operated platforms: the Claude API, Claude Platform on AWS, and Microsoft Foundry. Amazon Bedrock and Google Cloud set their own schedules, so if you run on a partner platform, none of the floors above bind you.
FAQ
Is Claude Opus 4.1 still usable through the API?
No. claude-opus-4-1-20250805 was retired on August 5, 2026. Anthropic's release notes state that all requests to this model now return an error. There is no silent fallback and no degraded response, so production code with that string hardcoded fails outright rather than quietly changing behaviour. Researchers can request continued access through Anthropic's External Researcher Access Program.
What is the exact retirement date for Claude Sonnet 4.5?
There is no exact date yet. The deprecations page lists claude-sonnet-4-5-20250929 as Active with a tentative retirement of "not sooner than September 29, 2026" and no deprecation date assigned. Because Anthropic commits to at least 60 days' notice and no notice had appeared as of August 6, 2026, that September 29 floor is already unreachable. The realistic earliest retirement is October 2026 or later.
How much warning does Anthropic give before retiring a model? At least 60 days for publicly released models, delivered by email to customers with active deployments and reflected in the documentation. In practice the recent notices have been almost exactly 60 days: Opus 4.1 got 61, Sonnet 4 and Opus 4 got 62, Haiku 3 got 60. Older models got much more, with Claude Opus 3 receiving 189 days. Plan on the minimum.
Should I migrate from Opus 4.1 to Opus 4.8 or Opus 5?
Anthropic's own docs disagree. The deprecation table names claude-opus-4-8; the August 5 release note recommends Claude Opus 5. They are priced identically. Opus 4.8 is the closer behavioural match, particularly if your code disables thinking at high effort levels, since Opus 5 returns a 400 error for disabled thinking at effort xhigh or max. Runway is close to a wash: Opus 5 buys about two extra months, a July 24, 2027 floor against May 28, 2027.
Do these retirement dates apply on Amazon Bedrock and Google Cloud? No. Anthropic states the dates apply to Anthropic-operated platforms only: the Claude API, Claude Platform on AWS, and Microsoft Foundry. Amazon Bedrock and Google Cloud are partner-operated and set their own retirement schedules, so a model's lifecycle status and dates can differ there. Check the Bedrock and Google Cloud model tables separately rather than assuming the Anthropic timeline carries over.
Do dateless model IDs like claude-opus-5 auto-upgrade?
No. Anthropic's models overview is explicit that starting with the 4.6 generation, model IDs use a dateless format that is still a pinned snapshot and not an evergreen pointer. claude-opus-5 will not silently become a newer model. It has a retirement floor of July 24, 2027 exactly like any dated ID. For models before the 4.6 generation, the alias column entries are convenience pointers that resolve to a dated ID.
Why does Anthropic retire models at all? Capacity, stated plainly in the docs: "Anthropic currently deprecates and retires models to ensure capacity for new model releases." The research post adds that cost and complexity of serving models for inference scales roughly linearly with the number served. It is not framed as a safety improvement. Anthropic separately lists the downsides, including lost research access and what it calls model welfare risks.
Are retired Claude model weights deleted? No. Anthropic has committed to preserving the weights of all publicly released models, and models deployed for significant internal use, for at minimum the lifetime of the company. It also produces and preserves a post-deployment report for each deprecated model, including a recorded interview with the model about its own deployment. Anthropic says it hopes to make past models publicly available again at some point.