Build1 publisher3 min readPublished
OpenAI's retirement list sets shutdown dates for about fifty snapshots, GPT-4 among them
OpenAI's deprecations page lists about fifty model snapshots, GPT-4 among them, with shutdown dates between September 24 and December 11, 2026. How each team's code fails on those dates depends on whether it pinned a snapshot or called an alias.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- A call to a retired OpenAI snapshot returns HTTP 404, with type invalid_request_error and code model_not_found.
- An alias such as gpt-4o resolves to whatever snapshot OpenAI currently maps it to, so a remap raises no error while prompts, parsers and evals drift.
- OpenAI promises at least six months' notice for generally available models, three months for specialised variants and as little as two weeks for previews.
- Anthropic retired nine Claude models in twelve months, from two Claude 3.5 Sonnet snapshots in October 2025 to Opus 4.1 in August.
- On Claude 4.7 and newer, a non-default temperature, top_p or top_k returns a 400 error, and Anthropic's Python SDK removed those parameters from v1.0.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Each call site needs an explicit choice between a dated migration on a pinned snapshot and an untracked model change behind an alias, and the alias holds up only where evals watch live traffic.
- constraint Moving to Claude 4.7 or newer means deleting sampling settings, so teams that used temperature to keep outputs reproducible lose that control in the same change as the model swap.
- exposure A Claude model called through Bedrock or Google Cloud can retire on a different day than on Anthropic's own platform, so a multi-cloud deployment has to track more than one published date.
- cost Building on an OpenAI preview model commits a team to migrating on as little as two weeks' warning, about one sprint.
The dev.to post that compiled the list shows both cases in four lines [8]:
```python # pinned: dies loudly on retirement day model = "gpt-4o-2024-05-13" # 404 model_not_found (shutdown Oct 23, 2026) # alias: upgrades silently, evals drift model = "gpt-4o" ```
The first line stops working on October 23 [8]. Earlier retirements show what the caller sees. Public issue trackers quote the message verbatim: "The model 'text-davinci-003' has been deprecated, learn more here: https://platform.openai.com/docs/deprecations" [17]. The link goes to a table. "This is the good failure: it is loud, it happens on a known date, and your monitoring sees it," the post's author wrote [18].
The second line has no date at all [7]. An alias is safe only if evals run against live traffic often enough to catch a model change as a shift in a metric. If they run at release time, the new model answers until the next release. In my context, a downstream parser depends on output shape and evals run per release. So I'd pin every call site and own the migration date. A team with continuous production evals could reasonably keep the alias.
Owning the date is manageable because the dates are public. The fifty-odd OpenAI shutdowns fall inside a 78-day stretch [1][1]. The September 28 batch retires the last GPT-3-era names [13], and the page lists 32 deprecation announcements since June 2023 [14]. A notice guarantees time, and the post notes that the successor is up to the vendor [12]. Sora's deprecation was announced March 24 with six months of notice, and its replacement column reads three dashes [12], the most candid cell on the page. Six months from March 24 is September 24, the first day of the window [2].
Anthropic adds one thing OpenAI does not [10]. Every active Claude model carries a "Tentative retirement date" from launch, set "Not sooner than" roughly a year out. For claude-opus-4-6 that date is February 5, 2027 [10]. That date can go in the adoption review on the day the model is chosen. Anthropic's promised floor is at least 60 days, and it has become the usual window: Opus 4.1 got 61 days, June 5 to August 5, 2026, and Sonnet 4 and Opus 4 got 62 [5]. Sonnet 3.7 got about four months and Opus 3 about six [5].
For a closed model behind an API, retirement is final, because retired means requests fail [3][16]. Anthropic has pledged to preserve the weights of all its publicly released models for as long as the company exists, and to interview each model before retiring it [15]. The pledge stops there: Anthropic makes no promise to act on what comes out of the interview, and none to publish any weights [15]. Open weights can outlive their host. Pirate Face mirrors about 669,000 Apache and MIT models from Hugging Face as torrents, and it can only save what was published [16].
What to watch
- Whether OpenAI names a replacement in the Sora row before its six months of notice run out.
- Which snapshot the gpt-4o alias resolves to after gpt-4o-2024-05-13 shuts down on October 23, 2026.
- Whether Bedrock and Google Cloud publish their own retirement dates for the Claude models Anthropic retires, and how far those dates diverge.