OpenAI removes gpt-4, o1, o3-mini and dozens more from the API on 23 October 2026
Over 25 model IDs shut down permanently on 23 October 2026. Here's what replaces each one and what it will cost you.
On 23 October 2026, OpenAI permanently shuts down over 25 model IDs across the API. Requests to any of them will start returning errors. There is no grace period, no automatic fallback, and no migration of fine-tuned weights. If your code still references any of the models below, you have a breaking change on the calendar.
What’s being shut down on 23 October 2026
The full list of models losing access on that date:
gpt-3.5-turbo(and all its dated snapshots)gpt-4/gpt-4-0613gpt-4-1106-previewgpt-4-turbo/gpt-4-turbo-2024-04-09gpt-4o-2024-05-13gpt-4.1-nanoo1/o1-2024-12-17o1-proo3-minio4-minigpt-image-1
Fine-tuned models built on these bases also shut down on the same date, including ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano-2025-04-14, ft-babbage-002, and ft-davinci-002. There is no automatic migration of fine-tuned weights to a new base model. If you have custom fine-tunes in production, you will need to retrain on a supported base.
What replaces each model
OpenAI’s official deprecations page lists a named replacement for every retired model ID. The table is not a single swap; three different models cover the range:
| Retiring model | Replacement |
|---|---|
gpt-3.5-turbo | gpt-5.6-terra |
gpt-4, gpt-4-1106-preview, gpt-4-turbo, gpt-4o-2024-05-13, o1, o1-pro, o3-mini | gpt-5.6-sol |
gpt-4.1-nano | gpt-5.6-luna |
o4-mini | gpt-5.6-terra |
gpt-image-1 | gpt-image-2 |
Fine-tuned models follow the same pattern: ft-gpt-3.5-turbo and ft-babbage-002 and ft-davinci-002 go to gpt-5.6-terra, ft-gpt-4 to gpt-5.6-sol, and ft-gpt-4.1-nano to gpt-5.6-luna.
What this means for your costs
This is where things get uncomfortable for teams running cost-optimised pipelines. At the list prices published when OpenAI announced these deprecations in April 2026, switching from some of the cheaper retiring models to their named replacements is a meaningful price increase:
gpt-3.5-turboto gpt-5.6-terra: roughly 5.7x the costo3-minito gpt-5.6-sol: roughly 6.3x the costo4-miniandgpt-4.1-nanoto their replacements: roughly 2.5x the cost
For teams that chose these models specifically because they were cheap, that is a real budget impact and worth modelling before the deadline.
The picture is more nuanced for the higher-end models. For gpt-4, gpt-4-turbo, and o1, moving to gpt-5.6-sol is actually cheaper at list price, not more expensive.
It is also worth noting that gpt-5.6-sol carries a promotional rate of $4 per million input tokens and $20 per million output tokens, available through at least 21 November 2026, versus the standard $5/$30 rate. GPT-6 Sol and Luna, the current generation at the time of writing, are priced at $2/$10 and $0.10/$0.50 respectively, which is substantially cheaper than gpt-5.6-sol for most workloads. If you are re-evaluating your model choice anyway, it is worth checking whether a current-generation model is a better fit before you simply drop in the named replacement.
The alias trap
One thing worth flagging: many of the retiring model names are aliases that point to specific dated snapshots. Calling gpt-4 in your code has been equivalent to calling gpt-4-0613. Calling o1 resolves to o1-2024-12-17. Both the alias and the snapshot it points to are being shut down. Checking your code for the alias is not enough; you should also audit for any hardcoded dated snapshot strings.
What else is changing around the same time
The October shutdown is the largest single wave, but it is not the only change happening in this period. The Evals platform becomes read-only on 31 October 2026 and shuts down 30 November 2026. The Assistants API shut down on 26 August 2026. A second deprecation wave follows on 11 December 2026, covering gpt-5 dated snapshots and o3-pro, and a third wave on 14 January 2027 takes down gpt-4o-2024-08-06, gpt-4o-mini, gpt-4.1, and gpt-4.1-mini. If you have migrations to plan, now is a good time to look at all three dates together rather than treating 23 October as an isolated event.
What you should do before 23 October 2026
Audit your model references. Search your codebase, configuration files, and any routing layers for every string in the shutdown list, including dated snapshot IDs, not just the short aliases.
Check your fine-tunes. If you have ft-gpt-3.5-turbo or ft-gpt-4 models serving production traffic, inference on them stops on 23 October. Plan your retraining timeline accordingly; fine-tuning a new model takes time to validate in production.
Model your cost delta. Before swapping to the named replacement, check current pricing across gpt-5.6-terra, gpt-5.6-sol, gpt-5.6-luna, and the GPT-6 generation. The named replacement is not always the cheapest current option.
Test in staging first. The replacement models are not drop-in compatible in every case. Response format, token counts, and latency characteristics differ. Give yourself time to catch regressions before the forced cutover.
OpenAI’s standard notification policy is at least six months’ notice for generally available models. The April 2026 announcement gave teams until October, which is the full window. If you have not started yet, the deadline is close enough to treat this as urgent.