OpenAI's own deprecations page lists October 23, 2026 as the date GPT-3.5-turbo, GPT-4, and a wave of other legacy model snapshots stop answering entirely -- the widest single cutoff the company has scheduled to date. That's not a downgrade. It's a disclosed retirement, announced months ahead, with a named replacement. The actual downgrade already has a documented case, and it's older: in 2023, Stanford and UC Berkeley researchers found GPT-4's accuracy on a basic "is this number prime?" check fall from 97.6% to 2.4% between two versions carrying the exact same model name, change announced nowhere. Two different problems produce the identical complaint -- "my AI got worse" -- and there's a fast, checkable way to tell which one is actually happening to you.
Two different things hide behind the same complaint
The first is real, measured model drift: the same product name producing different output over time, with no version bump a user would ever see. The Stanford/Berkeley study is still the clearest public evidence of it. Comparing GPT-3.5 and GPT-4 snapshots three months apart in 2023, the researchers found GPT-4 grew less willing to answer sensitive questions, made more formatting mistakes generating code, and, most sharply, went from 97.6% to 2.4% accuracy on the prime-number check above -- all under one unchanged model name. OpenAI's own account of a related, separate episode that December put the underlying problem plainly: after users reported ChatGPT giving shorter, lazier answers, the company said the model "hadn't been updated since November 11" and called chat-model training "not a clean industrial process." Read together, both point at the same fact: quality can move without anything telling you it moved.
The second is ordinary vendor housekeeping that looks identical from the outside if you never check the calendar: a scheduled, disclosed retirement with a named replacement and a real lead time. None of the three deprecation pages behind this guide read like a model card explaining what changed in an update -- they're lifecycle calendars, dates and replacement names, nothing about behavior. But a calendar is checkable, and a feeling isn't.
What each vendor actually promises, and where to check it
OpenAI, Anthropic, and Google Gemini all publish real deprecation schedules now. The guarantees underneath them differ enough to matter: Anthropic promises at least 60 days' notice before a publicly released Claude model retires, while Google's own gemini-2.0-flash and other 2.0-series models are already scheduled to retire June 1, 2026, with no fixed minimum notice period stated at all.
Model deprecation policy: OpenAI vs. Anthropic vs. Google Gemini
| OpenAI | Anthropic | Google Gemini | |
|---|---|---|---|
| Minimum notice for a released model | At least 6 months (GA); as little as 2 weeks for preview models | At least 60 days | No fixed minimum -- listed dates are 'earliest possible' |
| Where to check your exact model | developers.openai.com deprecations page | platform.claude.com deprecations page + Console usage export | ai.google.dev deprecations + release notes pages |
| A confirmed 2026 example | GPT-3.5-turbo and GPT-4 legacy snapshots retire October 23, 2026 | Opus 4.1 retired August 5, 2026 -- 61 days after notice | gemini-2.0-flash and variants retire June 1, 2026 |
That gap holds even within one vendor. An independent developer audit of Anthropic's own published numbers found the same Claude Opus 4.1 retirement carrying 61 days' notice on Anthropic's direct API and 184 days on AWS Bedrock -- same model, same retirement date, because the two platforms start the notice clock at a different moment. The platform you're actually calling can matter as much as which model you picked.
Check whether your AI chatbot actually got downgraded
- Check your API integration's model field, or the app's model-selector setting, for a dated snapshot name versus a bare 'latest' or default label.
- OpenAI, Anthropic, and Google each list every model's current lifecycle stage and retirement date on the pages in this guide's sources -- not a forum thread, the vendor's own documentation.
- Some products openly route free-tier or high-demand users to a lighter model under load -- a real, disclosed practice, and not the same problem as undisclosed drift.
- Save a fixed prompt and its answer, then re-run the same prompt later and compare against that saved baseline -- not against a general impression of last month.
None of this requires taking a vendor's word over your own experience. It requires checking a specific page instead of a general impression -- the same discipline our guide to verifying an AI benchmark claim applies to a company's own performance claims holds up just as well pointed at your own account. And the underlying tradeoff is the same one our guide to deciding between self-hosting and an API covers at a larger scale: an API user gets every improvement a vendor ships for free, and inherits every change for free too, with no more warning than that vendor's own deprecations page provides.
Three ways this gets misdiagnosed
The larger shift here is structural, not a conspiracy: three vendors are now retiring models on a real, published cadence, at a scale and specificity that wasn't public two years ago. That's progress -- a checkable calendar beats a vibe. Check it before concluding a model got worse. Often the honest answer is that it didn't change at all, and neither did you -- the model just kept doing exactly what an unpinned 'latest' alias always promised it might.
- A 2023 study caught GPT-4's math accuracy fall from 97.6% to 2.4% with no name change.
- OpenAI retires GPT-3.5-turbo and GPT-4 outright on October 23, 2026 -- its widest cutoff yet.
- OpenAI, Anthropic, and Google each publish deprecation dates, but on different minimum-notice schedules.
- Check whether your app calls a pinned snapshot or a moving 'latest' alias first.
- Caveat: a documented, scheduled retirement isn't the same problem as undisclosed quality drift.