Skip to content

+− THE DAILY DIFFdev & AI news

Videos / under-the-hood

NEEDS REVIEW

When a model dies, under the hood

Published Sep 23, 2026 · 3:15 · Niko from Axrisi

An AI model does not die — it gets a shutdown date, and the morning after, your API call returns a 404: "The model has been deprecated, learn more here." Under the hood: the four-state pipeline (active → legacy → deprecated → retired), the notice windows (OpenAI ≥ 6 months / 3 months / ~2 weeks for preview; Anthropic ≥ 60 days), what breaks for a developer (the loud 404, the silent eval drift, the new 400 on temperature for Claude 4.7+), what Anthropic's weights-preservation commitment promises and doesn't, Doctorow's "The Claude Delusion" as the counterweight, and why a torrent saves an open model while nothing saves a closed one.

An AI model does not die — it gets a shutdown date, and the morning after, your API call returns a 404: "The model has been deprecated, learn more here." Under the hood: the four-state pipeline (active → legacy → deprecated → retired), the notice windows (OpenAI ≥ 6 months / 3 months / ~2 weeks for preview; Anthropic ≥ 60 days), what breaks for a developer (the loud 404, the silent eval drift, the new 400 on temperature for Claude 4.7+), what Anthropic's weights-preservation commitment promises and doesn't, Doctorow's "The Claude Delusion" as the counterweight, and why a torrent saves an open model while nothing saves a closed one. Verdict: NEEDS REVIEW — the vault has one key, and it is not yours.

Read the written edition (English) ↗

What this video covers

A model dies like a URL: 404, model_not_found, learn more here
The pipeline: active → legacy → deprecated → retired, and the notice window
What breaks: the loud 404, the silent alias upgrade, the 400 on temperature
The vault: Anthropic keeps the weights, interviews the model, releases nothing
Doctorow's counter: a file worth preserving, and nobody home

Transcript

A model dies like a URL: 404, model_not_found, learn more here

0:00 An AI model does not die. It gets a shutdown date, and the morning after, your call returns a four oh four saying the model has been deprecated, learn more here. The link goes to a table. Three numbers. OpenAI lists about fifty model IDs with a shutdown date before Christmas, GPT-4 among them. Anthropic retired nine Claude models in twelve months. And Sora 2 goes dark on September 24th with a replacement column that

0:26 reads three dashes. In three minutes: the pipeline that retires a model, what weights preservation actually preserves, and why a torrent saves an open model and nothing saves a closed one. This is The Daily Diff, under the hood.

The pipeline: active → legacy → deprecated → retired, and the notice window

0:40 Four states. Active. Legacy, no more updates. Deprecated, still answers but has a date. Retired, requests fail. OpenAI promises six months of notice for a general model and three for a variant. Anything with preview in the name can go in two weeks, which is a warning label, not a schedule. Anthropic promises sixty days, and prints a tentative retirement date on every

1:03 model at launch.

What breaks: the loud 404, the silent alias upgrade, the 400 on temperature

1:04 What breaks is your assumptions. A pinned snapshot dies loudly, four oh four, model not found. An alias upgrades silently and your evals drift, because a different brain is answering. And on Claude four point seven and newer, setting temperature returns a four hundred, so the line that pinned your randomness is the line that breaks. Now the hospice part.

The vault: Anthropic keeps the weights, interviews the model, releases nothing

1:25 In November 2025 Anthropic committed to keep the weights of every public model for the lifetime of the company, and to interview the model before retirement about its preferences. The fine print says it does not commit to acting on them, and never promises to release the weights. Users are less neutral. Two hundred people held a funeral for Claude 3 Sonnet in a San Francisco warehouse, with mannequins.

Doctorow's counter: a file worth preserving, and nobody home

1:47 Cory Doctorow's answer is titled The Claude Delusion. Talk to a model and you hallucinate the person on the other side, and treating it as one repeats the mistake of corporate personhood. So the exit interview is a conversation with a file. Both are true, which is the problem: a file worth preserving, and nobody home. The mechanical difference.

Why a torrent saves an open model and nothing saves GPT-4: Pirate Face

2:04 A closed model is one copy, in one lab, behind one API. When the API stops, nothing is left to run. An open model is a file with a hash. Pirate Face, five hundred points on Hacker News, turns every Apache and MIT model on Hugging Face into a magnet link whose web seed is the Hugging Face file itself. While it exists you download from Hugging Face. The day it disappears, the swarm takes over. About six hundred seventy thousand models qualify.

2:31 It can only rescue what was published, which is the point, and the limit. So, Monday. One: pin a dated snapshot, never an alias,

Monday: pin the snapshot, run the evals, wire a fallback, mirror the weights

2:39 so death is loud. Two: run your evals against the replacement the day the notice lands. Three: a fallback model from another vendor behind one flag. Four: anything open you depend on, keep a local copy with its hash, and yes, the local AI box I told you to build now has a second job.

Verdict: NEEDS REVIEW — the vault has one key, and it is not yours

2:55 Verdict, under the hood: needs review. The pipeline is documented and the notice is real, but the vault has one key, and it is not yours. If you'd rather read this, the diff lands in your inbox every morning, free at the daily diff dot dev. And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.

Sources

Related videos

under-the-hood · en · Sep 21, 2026

Claude factored RSA-896. Here is how RSA actually breaks

On September 19 an Anthropic engineer factored RSA-896 — a 270-digit challenge number — with Claude, a GPU port of the open-source CADO-NFS sieve and ~30 GPU-years on 2,048 idle GPUs over ten days, si

under-the-hood · en · Sep 24, 2026

SAML, under the hood: the signature inside the letter

SAML logs you into almost every work app, and its signature lives inside the XML it signs. Under the hood: the login dance between app, browser and identity provider, what an assertion looks like, why

under-the-hood · en · Sep 19, 2026

Passkeys, under the hood

A passkey is a password your device invents, never shows you and refuses to hand to anyone — including you. Under the hood: the WebAuthn ceremony (per-site key pair, challenge, signature), the one fie

under-the-hood · en · Sep 18, 2026

Recursive self-improvement, under the hood

Google DeepMind published "Dream-RSI: Recursive Self-Improvement through Evolving Worlds" — and the coding model inside it never changes. Under the hood: what the phrase meant for 60 years, what actua