OpenAI retired GPT-4.5 from ChatGPT on June 26-27, automatically migrating every user and Custom GPT to GPT-5.5. The model that powered six months of content workflows is gone, and most marketing teams have no idea the transition happened.
The content brief template your team dialed in over a hundred iterations, the Custom GPT your agency built to match your client's brand voice, the research assistant that halved the time your analyst spent on competitor summaries, all of them ran on a different AI model this weekend than they did last week. OpenAI retired GPT-4.5 from ChatGPT on June 26-27, completing a 30-day migration that moved every ChatGPT user and every Custom GPT to GPT-5.5 automatically. The retirement was announced in OpenAI's May 28 release notes. Most teams that use ChatGPT for work never saw the notification.
The scope is worth being precise about. This is a change to the ChatGPT consumer and business interface, meaning the tool most marketing managers, content strategists, and account teams open in their browser to do actual work each day. Developers who pin API calls to specific model versions are unaffected for now. But anyone who opens ChatGPT, selects a Custom GPT, or uses the standard chat interface has been running GPT-5.5 since this weekend. Existing conversation threads auto-migrated. New sessions default to GPT-5.5. The model name surfaces in the interface if you know where to look. Most users are not looking.
What the upgrade actually means for output behavior
GPT-5.5 is measurably better than GPT-4.5 on most dimensions that matter for knowledge work. It reasons more reliably across multi-step tasks, follows complex instructions with fewer errors, and handles nuanced tone requests with more consistency. OpenAI framed the retirement as a straightforward upgrade, and by most benchmarks, they are right.
The complication is that "better on average" does not mean "identical in behavior." GPT-5.5 tends to produce more thorough responses at the same prompt length. It applies more caution when interpreting ambiguous instructions. It adds qualifications and caveats in contexts where GPT-4.5 would have produced a flatter, more direct output. For marketing teams that spent months calibrating prompts to generate specific output lengths, specific tonal registers, or specific structural formats, those calibrations are now running against a model that responds differently than what they were tuned against.
Subject lines dialed in for a specific character count need re-testing. Content calibrated to hit a specific reading level or brand voice register may be landing slightly more formal than expected. Any saved "sample output" used to brief freelancers or train reviewers represents GPT-4.5 behavior. Reviewers comparing new ChatGPT work against those samples are now working with two different models without knowing it.
Custom GPTs are where the drift goes unnoticed
The retirement also migrated every Custom GPT built on GPT-4.5 to GPT-5.5 as the underlying engine. For agencies and marketing teams that built specialized assistants, brand voice tools, campaign ideation generators, landing page drafters, the tool looks identical from the outside. The system prompt is unchanged. The name is unchanged. The output behavior is different.
This is the category of invisible drift that causes quiet problems over time. A brand voice Custom GPT that was validated against 200 GPT-4.5 outputs is now generating GPT-5.5 outputs, and the human reviewers who calibrated their judgment against the earlier samples have no signal that something shifted. The outputs are not worse. But the distribution of surprises is different, and nobody has recalibrated for it.
The broader pattern is worth naming directly. With GPT-4.5 now retired and GPT-o3 scheduled for retirement in August, OpenAI is completing the full transition to the GPT-5 family as the default baseline across ChatGPT. This is the third model retirement cycle in under 18 months. The underlying AI model powering most business ChatGPT usage has turned over roughly every four to six months. Any team that assumed their tool was stable has been running through silent model changes without knowing it. That assumption is not unusual. It is just not accurate.
The honest caveat
The risk from this transition is not decline. GPT-5.5 is a genuine upgrade and for most marketing tasks the outputs will be at least as good and often better. The risk is invisible drift: teams that built workflows on GPT-4.5 behavior, validated outputs against it, and set quality standards from those samples are now getting subtly different results without a signal that anything changed.
Teams that actively track output quality, run periodic spot-checks on their prompt libraries, and treat AI calibration as an ongoing practice rather than a one-time setup will absorb this transition without disruption. Teams that treat the AI as a fixed tool will notice something feels slightly off without being able to explain why, usually at the moment a client or reviewer flags it.
One additional note on scope: the API is unaffected. Developers who specifically need GPT-4.5 behavior for comparison purposes or to maintain parity with previously published benchmark outputs can still access it through the API. The retirement applies only to the ChatGPT consumer and business interfaces, not to developer API endpoints.
The closing observation
There is a version of the GPT-4.5 retirement that most teams experience as nothing. They open ChatGPT, it produces output, nobody notices a model changed. That seamlessness is either the best case scenario or the quietly worst one, depending on whether anyone is actually watching. The teams that do not monitor AI output behavior discover drift at the moment a client, customer, or internal reviewer flags something that felt different but could not be explained. Model retirements are not emergencies. They are scheduled calibration moments. The difference between teams that handle them smoothly and teams that do not is simply whether they knew to look.