Saturday means a shorter model menu in ChatGPT. OpenAI is pulling GPT-4.5 from its model selector, and if you have Custom GPTs pinned to that model, you have until Friday to test what changes. In August, o3 follows. Neither removal is dramatic, but both deserve a 30-minute check before Monday catches you off guard.
What exactly is leaving ChatGPT?
GPT-4.5 disappears from the ChatGPT model picker on June 27. Paid users on Plus, Team, and Enterprise have been able to select it manually since early 2025, as an alternative to the default GPT-5 line-up. After Saturday, that option is gone.
Next in line: o3. The reasoning model exits ChatGPT on August 26, with a 90-day window instead of the 30 days given to GPT-4.5. That difference is meaningful. OpenAI's longer runway for o3 signals that the model sits deeper in user workflows, and the company is giving teams more time to migrate reasoning-heavy tasks to GPT-5.5.
Worth noting: in the OpenAI API, GPT-4.5 was already removed on July 14, 2025. Developers using the API will notice nothing from this change. This is purely a ChatGPT interface decision, the website and the app. That gap makes it all the more striking that it took OpenAI a full year to remove the same model from its consumer product.
| Model | Leaves ChatGPT | API status | Successor | Sunset window |
|---|---|---|---|---|
| GPT-4.5 | June 27, 2026 | Already removed (July 2025) | GPT-5.5 | 30 days |
| o3 | August 26, 2026 | Until December 11, 2026 | GPT-5.5 | 90 days |
| o3-pro | August 26, 2026 | Until December 11, 2026 | GPT-5.5-pro | 90 days |
What made GPT-4.5 worth keeping?
GPT-4.5 was the largest model in the GPT-4 generation. Slower and more expensive than GPT-4o, but noticeably better at nuanced writing, complex reasoning, and sustained long conversations. It was the model you reached for when you needed careful, not fast.
Think of it as the difference between a quick email and a thoughtful memo. GPT-4o wrote the email. GPT-4.5 wrote the memo. That choice disappears Saturday, but GPT-5.5 outperforms GPT-4.5 across nearly every benchmark. OpenAI's newest flagship scores 82.7% on Terminal-Bench 2.0, more than 13 points ahead of Claude Opus 4.7 (69.4%), and significantly higher than anything GPT-4.5 could manage on the same benchmark. You are losing an option, but gaining a stronger one.
Still, some teams swear by GPT-4.5's specific behavior: the way it phrases things, response length, the style of its reasoning. These subtleties don't show up in benchmark tables, but they show up every day in the output your team has learned to rely on. Switching models is like handing work to a different writer. Even a technically better one produces different output.
Hallucination rates were another consideration. Data from Vectara showed GPT-4.5 performed comparably to other GPT-4-generation models on accuracy, while GPT-5.5 shows meaningful improvement. That is a genuine advantage of the upgrade, particularly for business use cases where factual precision matters.
What happens to your Custom GPTs?
Custom GPTs set to GPT-4.5 will automatically switch to the default model after the deadline. Right now, that means GPT-5.5.
For context: that sounds harmless, and usually it is. GPT-5.5 scores higher on virtually every benchmark: coding, reasoning, multilingual tasks. But "better" is not the same as "identical." If your Custom GPT is tuned to GPT-4.5's specific behavior, a writing style, a way of structuring answers, a carefully calibrated system prompt, the switch can introduce subtle shifts in output.
Here's the thing: subtle shifts are exactly the kind you miss until they have already reached a client. If your team runs a Custom GPT that summarizes proposals in a fixed format, the tone might drift slightly more formal, the structure might reorder. Not wrong, but different. And "different" is what you don't want to discover after the document has gone out.
The impact depends on how you actually use ChatGPT. Teams that always used the default model and never manually selected GPT-4.5 will notice nothing. Teams that deliberately chose GPT-4.5 for specific tasks will feel the switch the most.
The same logic applies to o3-dependent workflows. If your team uses o3 for complex analysis or reasoning chains, start testing on GPT-5.5 now. You have until August 26, but the migration is easier when there is no urgency.
How do you check your workflows before Saturday?
This takes about half an hour:
- Go to chatgpt.com/gpts/mine. That is your full Custom GPT inventory.
- Open each Custom GPT and check the assigned model. If it shows GPT-4.5, it switches automatically on Saturday.
- Manually switch to GPT-5.5 now. This lets you test how output changes before the deadline forces it.
- Run your standard prompts. Feed each GPT the same inputs your team uses daily. Compare outputs: watch for changes in tone, structure, and content.
- Save your system prompts. If GPT-5.5 output does not satisfy you, a prompt adjustment is often enough. A system prompt tuned for GPT-4.5 may need rephrasing for GPT-5.5.
Using o3 for reasoning tasks? You have until August 26, but the same testing logic applies. Start now while the comparison is still calm.
What does this cost you?
Nothing extra. Your ChatGPT Plus subscription stays at the same price. You are losing a model from the menu, but GPT-5.5, its replacement, is stronger in most use cases. The subscription does not lose value.
There is one exception worth flagging for businesses that also use the OpenAI API alongside ChatGPT. The GPT-5.5 token pricing differs from o3's. If you have o3 embedded in API integrations, check the per-token cost before the API deprecation arrives in December. According to TheAIDaily's ChatGPT statistics, ChatGPT now reaches more than 900 million weekly active users globally, holding a 61.8% share of the AI search tool market. The vast majority of those users will never notice this change: most use the default model and never selected GPT-4.5 manually. Teams that chose deliberately are the ones who need to act.
What are all AI providers doing right now?
This is not unique to OpenAI. Every major AI provider is consolidating its model lineup at the same time.
Anthropic retired Claude Sonnet 4 and Opus 4 (the April 2025 versions) earlier this month, replaced by Claude Sonnet 4.6 and Opus 4.8. Google replaced Gemini 2.5 Flash with the faster Gemini 3.5 Flash. The pattern is identical across the industry: fewer models, each better. The performance gap between the best and the tenth-best model has narrowed to 5.4 percentage points, according to the Stanford HAI AI Index Report 2026. The models are converging.
Think of it this way: your favorite restaurant just cut half the menu and promised that every remaining dish is better. That's probably true. But you would still want to taste it before guests arrive.
For you as a user, this creates two dynamics at once. Fewer models to choose from makes everyday decisions simpler: a three-option picker beats a fifteen-option picker. But it also means fewer fallback options. If GPT-5.5 has an outage, there is no GPT-4.5 anymore.
That is the strategic point teams should take from this. Businesses that build their entire AI workflow around a single provider and a single model are concentrating risk. Whether you use built-in AI or a standalone subscription, diversification pays off. Having Claude or Gemini ready as a backup is not pessimism. It is the kind of practical hedge that keeps operations running when a single model goes down.
What is OpenAI actually saying about this?
Remarkably little. OpenAI published the model removal via a Help Center article and its API deprecations page. No major blog post, no announcement. A quiet cleanup.
That fits the strategy. OpenAI wants you to stop thinking about which model you are using. The vision is GPT-5.x for everything: email, code, analysis, customer service. Model selection should be a developer problem, not a user problem. That is at least the idea.
In practice, plenty of teams chose specific models deliberately. A silent removal is less helpful than a clear migration guide. The clock is running either way.
What should you do next?
The immediate action is clear: check your Custom GPTs at chatgpt.com/gpts/mine, test them on GPT-5.5 today, and save your system prompts. That 30 minutes now saves you a Monday surprise.
The broader point is strategic. AI providers are consolidating their model rosters at an accelerating pace. The model landscape your team knew last month may look different next month. Actively tracking what AI costs your team and testing regularly means a model retirement does not catch you off guard, because this will keep happening.
Three things to do this week:
- Check your Custom GPTs at chatgpt.com/gpts/mine and manually switch to GPT-5.5.
- Test the output of your most-used workflows on the new model.
- Consider a second AI provider as a backup. Claude, Gemini, and Copilot are mature enough to step in when needed, not because GPT-5.5 is unreliable, but because dependency on a single provider is always a risk worth managing.
