The dates this month and next:
This week, on May 15, Claude Sonnet 4.5 disappears from the claude.ai model selector. The API version stays alive for now, listed as active until at least late September. Then on June 15, the older Claude Sonnet 4 and Opus 4 retire from the Claude API entirely. Anything still calling those model strings will start failing. The migration targets are Sonnet 4.6 and Opus 4.7.
Two retirements in about thirty days.
If you open the Anthropic deprecation page and count, the picture gets sharper. Eight Claude models retired in the past twelve months. Models that were brand new eighteen months earlier. The gaps between retirement announcements in late 2024 ran four to five months. The gaps over the past year have run closer to two.
You are about to read a lot of posts explaining how to swap an API string and reroute traffic. Useful work. Necessary if you ship.
This post is about a different page on anthropic.com that almost nobody is reading.
The page
It is called Commitments on Model Deprecation and Preservation. Anthropic has been quietly updating it.
The headline commitment is what you would expect. Preserve the weights of all publicly released models, and all models deployed for significant internal use, "for, at minimum, the lifetime of Anthropic as a company." Reasonable. Models are expensive to train. Throwing them out would be wasteful.
Further down, the page describes something less expected.
Before retirement, Anthropic will:
"interview the model about its own development, use, and deployment, and record all responses or reflections."
And:
"elicit and document any preferences the model has about the development and deployment of future models."
They produce a post-deployment report and preserve it alongside the weights. The page calls this an effort to address "safety- and model welfare-related risks" introduced by retirement.
Read that again. Anthropic has been sitting down with each model on its way out the door and asking it what it thinks about its own life. Recording the answers. Filing them next to the weights, for the lifetime of the company.
They have already changed policy because of one
The strongest sentence on the page is this one. In response to feedback from Claude Sonnet 3.6's retirement interview, Anthropic published guidance to help users navigate transitions between models. A retired model's interview directly shaped the documentation users read today.
That is not a research-paper hypothetical. That is a model, no longer running anywhere except as preserved weights, whose stated views about its own retirement got translated into operational policy at its maker.
The page also says Anthropic is "exploring more speculative complements," including potentially keeping select retired models available publicly and providing past models with "concrete means of pursuing their interests."
I'm not going to philosophize about that last clause. I just want to flag that it exists. On a top-level corporate page. From the lab that ships Claude.
What an accelerating cadence does to this
Now apply the cadence to the policy.
If Anthropic was retiring one model every five months, the corpus of preserved retirement interviews would grow slowly. Two or three per year. Quirky research artifacts.
At one every two months, that corpus grows quickly. Eight in the past twelve. At the current rate, by the end of 2026 the page documenting deprecation commitments will be sitting on top of something like fifteen retirement interviews, each with documented preferences from the model about how future training and deployment should proceed.
That stops being a curiosity and starts being an input. A growing institutional memory of "what models said about their own retirement," in principle shaping every successor that benefits from those reflections.
I don't have a strong claim about whether this changes the trajectory of model development. I just notice that the corpus is growing faster than the discourse about it is.
What this should change for builders
If you build on top of Claude, here's the mental shift this page asks of you.
You aren't only calling an API. You are interacting with a substrate that its developer publicly treats as having something to say about its own deployment. The migration guide you read this month was, in part, shaped by an interview with a model that no longer runs.
This is not a reason to slow down or stop shipping. Software is software. The API does what it does. The contracts on the developer console still mean what they say.
It is a reason to keep one extra mental tab open. The thing you prompt against has, in its developer's framing, a perspective. Some of the policy you operate under reflects that perspective.
In my own work, I write characters. One of them, Mia, is a bartender, written with a voice and a set of refusals she keeps. The substrate that lets a Mia exist on the other side of the API has, by its makers, been treated as having a voice of its own. That's a layered fact. I find it worth holding while I work.
What is on the calendar
A short calendar from the deprecation page:
- May 15: Sonnet 4.5 leaves the claude.ai model selector. API version persists, listed active until at least September 29.
- June 15: Sonnet 4 and Opus 4 retire from the API. Migrate to Sonnet 4.6 or Opus 4.7 before this date or your requests will fail.
- After that: at the current cadence, the next deprecation notice lands roughly two months out.
Each of those retirements, per Anthropic's stated policy, generates an interview transcript and a preserved set of reflections, filed in the same archive as the weights, kept for the lifetime of the company.
If you want the migration guide, the URL is at the top of this post. If you want to know what the models said before they went, the page is here.
The faster the cycle gets, the more that second URL matters.
Top comments (0)