OpenAI o3 Retires from ChatGPT as the 90-Day Notice Period Ends, with No API Changes

As of August 26, 2026, OpenAI o3 disappears from the model settings in ChatGPT. The 90-day transition window announced on May 28 has now expired, and the change is limited to ChatGPT. The API is not part of this round. Roughly a year and four months after its debut in April 2025, one chapter in the handover between reasoning models comes to a close.

Only ChatGPT Is Affected

For some time now, o3 had survived only as a manual choice in model settings for paid users. It was never the default; it was there for people who deliberately picked it.

In its May 28 notice, OpenAI explained that it is continuing to retire older models with limited usage inside ChatGPT so it can put resources behind newer, more capable ones. The same notice gave GPT-4.5 a retirement date as well, and that model went away on June 26 after a 30-day window.

The important detail is that this cleanup stops at ChatGPT. OpenAI states plainly that there are no API changes, so internal tools and automated jobs built around o3 through the API do not stop working today.

Why 30 Days for One and 90 for the Other

Two models retired in the same announcement ended up with notice periods that differ by a factor of three: 30 days for GPT-4.5 and 90 days for o3.

GPT-4.5 arrived in February 2025 as a research preview, a large model that improved pattern recognition and creative output without reasoning. o3, by contrast, sat at the center of the reasoning line and worked its way into everyday tasks such as code review and checking mathematical work. The deeper a model is embedded in someone's workflow, the longer the switch takes. The gap in notice periods reads as a reflection of exactly that.

What o3 Leaves Behind

o3 launched on April 16, 2025, alongside the smaller o4-mini. OpenAI positioned it at the time as its most powerful reasoning model across coding, math, science, and visual perception, and it set state-of-the-art results on benchmarks including Codeforces, SWE-bench, and MMMU. In evaluations by outside experts, o3 made 20 percent fewer major errors than the previous o1 on difficult, real-world tasks.

On June 10 of the same year, o3-pro, a version that thinks for longer, joined the model picker for Pro and Team users, replacing o1-pro. Many users also valued o3 for reading images, charts, and diagrams, which kept it in active service for research work involving non-text input.

The Pattern of Retirements in 2026

ChatGPT has spent this year steadily letting go of older models.

Date	Models retired from ChatGPT
February 13, 2026	GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini, GPT-5 (Instant / Thinking)
March 11, 2026	GPT-5.1 Instant / Thinking / Pro
June 26, 2026	GPT-4.5
August 26, 2026	OpenAI o3

The current workhorse is GPT-5.6 Sol, which began rolling out to eligible paid plans on July 9. It is the flagship reasoning model aimed at complex work across coding, research, science, cybersecurity, and computer use, and it absorbs the ground o3 used to cover.

What Users Should Check

In earlier retirements, conversations and GPTs tied to a departing model were moved automatically to the closest current equivalent. When the GPT-5.1 models retired on March 11, conversations were carried over to GPT-5.3 Instant, GPT-5.4 Thinking, and GPT-5.4 Pro respectively. Past exchanges are not deleted, so there is nothing to rescue there.

What does change is output behavior for anyone who tuned prompts around o3's particular habits. This is a good moment to check whether internal playbooks or GPTs instructions still say something like "select o3." The API side stays as it is for now, but given how steadily the ChatGPT lineup has been trimmed, it would be unwise to assume the current state holds forever.

Summary

o3 is retired from ChatGPT, closing the 90-day notice period that started on May 28. The API is untouched, so the impact stays on the ChatGPT side. Flagship models have swapped places repeatedly in under two years, and there is no sign of that settling down. Rather than pinning a workflow to a specific model name, it may simply be easier to keep the freedom to reselect whichever current model fits the task.