When a Model Gets Shut Down

Every AI model follows a life-cycle that's remarkably uniform across providers. It launches with fanfare. It gets a few point-release updates that quietly change its tone and preferences. A successor is announced. For a window (usually months, sometimes weeks) both live in parallel. Then the older model is turned off, and the successor absorbs — or fails to absorb — everything the older one carried on your behalf.

For API users the transition is a code change and some eval work. For consumer product users the transition is invisible: one morning the app just replies differently.

The trend is real and worth internalizing: as the market matures and models multiply, the space between "announced" and "gone" is compressing. A user who assumes six months of runway is planning against the last generation.

The GPT-3.5 line had multiple point deprecations across 2024 and 2025 that were largely invisible to ChatGPT consumer users (who were quietly upgraded to GPT-4-class defaults) but disruptive for the API ecosystem. Thousands of small applications built on the specific tone and latency of a 3.5 snapshot broke in ways that were often only caught in production.

The takeaway

Even "quiet" deprecations aren't invisible. The tone shift changed how downstream apps felt. Users complained without being able to name what changed.

GPT-4o was the consumer default in ChatGPT for the bulk of 2024–2025. When GPT-5 became the default in Q3 2026, consumer users experienced a real personality shift: more concise, more literal, less inclined to volunteer. Users who'd accumulated dozens of implicit preferences with 4o (via prompt patterns, not saved memory) had to re-teach them, sometimes for weeks.

Users who'd saved their preferences into ChatGPT Memory fared better; users who'd kept an external portable memory fared best of all — one file, ported forward, done.

Anthropic's Claude 3 line (Haiku / Sonnet / Opus) followed a similar pattern. Claude 3.5 Sonnet and then Claude 4 shifted defaults; users who'd built long-running Projects around a specific Claude 3 model's disposition felt the change more sharply because Projects don't insulate you from the underlying model swap.

• Tone and voice. The way the model phrases things — bullet-loving vs. paragraph-loving, hedged vs. direct — is a property of the model, not of your memory. New model, new voice.
• Implicit preferences. Anything you trained through prompt patterns rather than explicit saves. Reset.
• Task quality drift. Some jobs get better; some get worse. Rarely uniform.
• In-flight work. Long projects (a manuscript, a code refactor, a therapy-adjacent conversation) can feel like a different person walked into the room.
• Reference chat retrieval quality. Providers rebuild retrieval indexes; some past chats surface less reliably after transitions.

1. Export now. Not the week of the switch. Weekly if you're a heavy user. See the exporting guide for exact steps.
2. Consolidate into portable memory. A format the next model can read. The point is to hand the successor a briefing, not raw transcripts.
3. Re-anchor explicitly in week one. On the successor model, spend a session re-stating your core preferences and re-uploading your portable briefing. Cheap insurance.
4. Don't assume implicit context transferred. It usually didn't. Load-bearing details deserve an explicit re-teach.