OpenAI Retires GPT-4 and GPT-3.5 for Good on October 23. Here's Every Model on the List
That's a real deadline, not a soft warning. Every GPT-3.5 and GPT-4 snapshot OpenAI still ran in production goes away at once, alongside its first image-generation model and four reasoning models from the o1/o3/o4 line.
The models most likely to actually break something are the ones nobody's looked at in a while: a Custom GPT built two years ago pointed at a specific snapshot, a production backend with a model name hardcoded in a config file, or a fine-tune nobody's retrained since the base model it sits on was current.
What's actually on the list
| Retiring October 23 | What it was | OpenAI's suggested move |
|---|---|---|
| gpt-3.5-turbo-0125 | last stable GPT-3.5 snapshot | gpt-5.6-terra |
| gpt-4-0613 / gpt-4-1106-preview / gpt-4-turbo | original GPT-4 line and its Turbo snapshots | gpt-5.6-sol |
| gpt-4.1-nano | smallest GPT-4.1 variant | gpt-5.6-luna |
| gpt-4o-2024-05-13 | first GPT-4o snapshot | a current GPT-4o or GPT-5.6-line model |
| gpt-image-1 | first-generation image model | gpt-image-2 |
| o1-2024-12-17, o1-pro-2025-03-19, o3-mini-2025-01-31, o4-mini-2025-04-16 | reasoning-tier models | current o-series or GPT-5.6 reasoning tier |
Replacement mappings are per OpenAI's deprecations page; check it directly before migrating, since context windows, pricing, and behavior differ from the models they replace.
Who this actually hits
- ✦Anyone with a fine-tune built on a retiring base model. Fine-tunes don't carry over automatically; you have to retrain on the recommended successor before the cutoff.
- ✦Older Custom GPTs and GPT Builder projects that still point at a specific legacy snapshot rather than a current default.
- ✦Production code with a model name hardcoded in an environment variable or config file that nobody's touched since launch.
- ✦Anyone using gpt-image-1 directly for a live feature; gpt-image-2 has different pricing and output behavior, not just a new name.
Before October 23
- 1Search your codebase for the exact model strings above. A hardcoded model name is the single most common way this breaks something in production.
- 2Test the suggested replacement in staging first. The GPT-5.6 line has different context limits and pricing from the models it's replacing; a drop-in swap can still change behavior.
- 3Retrain any fine-tune built on a retiring base now. Fine-tuned models don't migrate themselves, and retraining takes time you won't have on October 22.
- 4Check any Custom GPT you built or rely on. Its model setting may still point at a snapshot that's about to disappear.
When exactly do GPT-4 and GPT-3.5 shut down?
October 23, 2026, per OpenAI's own deprecation notice. All remaining GPT-3.5 and GPT-4 snapshots, plus gpt-image-1 and four o-series reasoning models, lose API access that day.
What happens if my app still calls a retired model?
The API call fails with an error. OpenAI doesn't automatically redirect requests to a newer model, so any code that hardcodes one of the retiring model names needs to be updated before October 23.
Does this affect ChatGPT, or only the API?
It's an API-level deprecation. It mainly affects developers calling these model IDs directly, and Custom GPTs or GPT Builder projects configured to use one of the retiring snapshots.
What should I replace gpt-4-turbo or gpt-3.5-turbo-0125 with?
OpenAI's deprecations page maps gpt-4-turbo and its related GPT-4 snapshots to gpt-5.6-sol, and gpt-3.5-turbo-0125 to gpt-5.6-terra. Test either in staging before switching production traffic over.
Do my fine-tuned models migrate automatically?
No. A fine-tune built on one of the retiring base models has to be retrained against the recommended successor; it does not carry over on its own.
Try the tools ✦
Free browser tools that never upload your files.