Summary: OpenAI announced on January 29, 2026 that GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini would leave ChatGPT on February 13, while GPT-5 Instant and GPT-5 Thinking would follow the same timetable. CNBC reported from OpenAI figures that only about 0.1% of users selected GPT-4o daily, but the change still touches a loyal long-tail after previous backlash. The API remains unaffected in this wave.
OpenAI picked January 29, 2026 to lay out the final schedule on its official blog: GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini come off ChatGPT on February 13, and the already-announced GPT-5 Instant and GPT-5 Thinking keep to that same timetable. Per CNBC, drawing on figures OpenAI provided, daily selection of GPT-4o has fallen to about 0.1% of users, while nearly everyone else settled on GPT-5.2 long ago. The API sits outside this retirement wave.
The story earns attention less because one more model is heading for the exit, and more because ChatGPT’s model menu is now formally narrowing to a single generation. Product teams get a real-time case study in how a supplier closes out long-tail model preferences: a devoted user base with deep emotional ties to an old model’s personality, up against an allocation logic that counts usage percentages.
The List, and the Day It Takes Effect
The announcement’s key points:
- Models leaving: GPT-4o (launched May 2024), GPT-4.1, GPT-4.1 mini, and o4-mini, along with the previously announced GPT-5 Instant and GPT-5 Thinking
- Effective date: February 13, 2026, when ChatGPT drops them
- Usage: by OpenAI’s count, only about 0.1% of users choose GPT-4o on a given day
- API: officially untouched — code with these model IDs hardcoded needs no change this time
The Case OpenAI Makes for Retiring Now
The justification rests on GPT-5’s recent gains in “personality, customization, and creative ideation” — far enough along, in OpenAI’s judgment, for GPT-4o to hand over without risk. The post is candid about the cost: “We know that losing access to GPT-4o will feel frustrating for some users, and we didn’t make this decision lightly,” and then “Retiring models is never easy, but it allows us to focus on improving the models most people use today.”
Everything turns on that 0.1%. Scaled against ChatGPT’s user base, the share still adds up to a large crowd, with a heavy tilt toward paying subscribers — GPT-4o made its name on a warm conversational feel, and part of the paying crowd stayed fiercely loyal for exactly that reason. OpenAI’s ledger, though, says the cost and complexity of running several legacy models now outweigh what serving that group brings in. Trimming model SKUs is the collective opening move of 2026 across frontier suppliers, and nearly all of them are pruning in the same direction (see our 2026 opening outlook).
This Wave Leaves the API Alone
CNBC’s reporting answers the developer question outright: the API is not affected. A product that rides on GPT-4o’s current behavior faces no forced migration in the near term. The real change lands in the experience layer — once the ChatGPT-side option is gone, admirers of the old personality have nowhere to point that preference, and how well they accept the new one will surface directly in feedback and retention numbers. And if “GPT-4o” still shows up in your marketing copy as a selling point, it is time to rewrite it.
The August 2025 Backlash and the “Plenty of Notice” Promise
The emotional pull of GPT-4o has a track record of making itself heard. When GPT-5 launched in August 2025, OpenAI cut off GPT-4o access entirely for a stretch, paying users pushed back hard, the company restored access within days, and CEO Sam Altman went on X to guarantee that any future retirement would come with “plenty of notice.” This round ran from the January 29 announcement to the February 13 effective date — roughly two weeks of lead time. It satisfies the pledge at the bare minimum, and it shows OpenAI has learned: retiring a model is now a product event that demands PR management, not a line item in backend maintenance.
Three Takeaways for Product and Engineering Teams
Three conclusions carry over. One: user attachment to a model’s personality is a genuine long-tail need, so decisions cannot rest on average usage alone — 0.1% of daily actives at ChatGPT’s scale is enough to set off a news cycle of its own. Two: suppliers narrowing their model menus is now the standard playbook, so product architecture should stand on a stable model family plus an abstraction layer, never on one particular legacy version. Three: the communication rhythm around a retirement has itself become a product capability — the length of the notice period and whether the replacement’s personality matches will both be inspected under a magnifying glass by the users most inclined to complain in public.