OpenAI Shuts Off o4-mini and GPT-3.5 Today, and Its Named Replacement Costs 8x More
OpenAI cuts API access to o4-mini, o3-mini, o1, GPT-3.5, GPT-4 and their fine-tuned derivatives on October 23. The replacements it names charge 8x more output than GPT-3.5 and 2.7x more than o4-mini, while the cheaper GPT-6 Luna is absent from the table.
- Calls to o4-mini, o3-mini, GPT-3.5 and GPT-4 start failing today, fine-tuned derivatives included.
- Following the official replacement table costs 8x more output for GPT-3.5 and 2.7x for o4-mini.
- GPT-6 Luna, one ninth of o4-mini's output price, is not on that table.
Requests to o4-mini, o3-mini, o1, gpt-3.5-turbo and gpt-4 start failing today. The date OpenAI wrote into its deprecations page back on April 22 is October 23, and the fine-tuned models built on those bases disappear the same day.
The page names a replacement for every model it retires. Put that table next to the pricing page and the arithmetic stops working: the cheaper the model you were running, the more the move costs you. Meanwhile the cheapest model OpenAI shipped in September is not in the table at all.
What goes dark today
The shutdown list lives under the entry "2026-04-22: Legacy GPT model snapshots." Dated snapshots and the floating aliases that point at them go together.
- Reasoning models:
o1,o1-pro,o3-mini,o4-mini - Older chat models:
gpt-3.5-turbo,gpt-4,gpt-4-turbo,gpt-4-1106-preview,gpt-4o-2024-05-13,gpt-4.1-nano - Image model:
gpt-image-1 - Fine-tuned models:
ft-o4-mini,ft-gpt-3.5-turbo,ft-gpt-4,ft-gpt-4.1-nano,ft-babbage-002,ft-davinci-002
What is absent from today's list matters just as much. Only the May 13, 2024 snapshot of gpt-4o goes; the gpt-4o alias itself, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini survive today, though the same page schedules all four for January 14, 2027. Full o3 and the first GPT-5 snapshots retire separately on December 11.
Nothing reroutes for you. OpenAI's notice does not say what a call to a dead model returns, or what time of day the switch flips. During the first wave on July 23, forum user _j reported in OpenAI's announcement thread that retired models stayed listed under /v1/models while actual calls came back with "deprecated" or "not found" errors. Health checks that confirm a model exists by listing it will not catch this.
Follow the table and the cheap models get expensive
Most of the named replacements are GPT-5.6 models shipped in July. Lined up against the pricing page at standard tier, per million tokens:
| Retired model (in/out) | Named replacement (in/out) | Output price change |
|---|---|---|
| gpt-3.5-turbo ($0.50 / $1.50) | gpt-5.6-terra ($2 / $12) | 8x |
| o3-mini ($1.10 / $4.40) | gpt-5.6-sol ($4 / $20, promotional) | 4.5x |
| gpt-4.1-nano ($0.10 / $0.40) | gpt-5.6-luna ($0.20 / $1.20) | 3x |
| o4-mini ($1.10 / $4.40) | gpt-5.6-terra ($2 / $12) | 2.7x |
| gpt-4o-2024-05-13 ($5 / $15) | gpt-5.6-sol ($4 / $20) | 1.3x |
| o1 ($15 / $60) | gpt-5.6-sol ($4 / $20) | one third |
| gpt-4 ($30 / $60) | gpt-5.6-sol ($4 / $20) | one third |
Teams on the expensive legacy models come out ahead. Moving from o1 or gpt-4 to GPT-5.6 Sol cuts output to a third. The penalty lands on everyone who picked a model because it was cheap. A service still routing bulk classification or summarization through gpt-3.5-turbo goes from $1.50 to $12 per million output tokens the moment it follows the table to Terra.
The GPT-5.6 Sol numbers carry a condition. The pricing page says only that Sol's promotional rate holds "at least until November 21, 2026." When OpenAI reshuffled prices on July 30, Sol's list price stayed at $5 input and $30 output. If the promotion lapses, an o3-mini workload moved to Sol goes from $4.40 to $30 per million output tokens, roughly 6.8x.
The cheapest option is a model OpenAI left off the table
OpenAI's own newest small model undercuts everything in that column. GPT-6 Luna, released September 22, costs $0.10 per million input tokens and $0.50 output. Against o4-mini that is one eleventh the input and roughly one ninth the output. Against GPT-5.6 Terra, the model OpenAI actually names as o4-mini's successor, output is one twenty-fourth the price. The deprecations page was written in April and never updated for the GPT-6 family.
Price is not the only direction it wins. On the Artificial Analysis Intelligence Index (v4.3.2), every replacement candidate scores above the models being retired.
Two caveats sit on those numbers.
- The legacy scores are estimates. Artificial Analysis marks the
o4-mini,o3-miniandgpt-4.1-nano(8 points) pages as awaiting independent evaluation. - The new scores come from maximum reasoning effort. Raising effort means more thinking tokens before an answer, so a low token price does not settle cost per request. Artificial Analysis measured $0.07 per index task for GPT-6 Luna against $0.18 for GPT-5.6 Luna.
For an o4-mini slot, that makes GPT-6 Luna worth testing before Terra. Terra leads by four index points and charges 24 times as much for output. Luna does get more expensive past its long-context threshold: above 272K input tokens it moves to $0.20 input and $0.75 output.
Who this hits, and why Azure users have more time
This is an API-only event. Nothing changes for people using ChatGPT, where o4-mini was already removed on February 13. None of the replacement models are region gated: an API key in any country OpenAI serves calls GPT-6 Luna, GPT-5.6 Terra and GPT-5.6 Sol at the prices above. The one regional wrinkle is pricing rather than access. Regional processing endpoints add a 10% uplift on models released on or after March 5, 2026, which covers every replacement named here. The pricing page does not list which regions are offered, pointing instead to OpenAI's data guide, so confirm your jurisdiction there before assuming a local processing option exists.
If you reach the same models through Microsoft Azure OpenAI, a separate retirement schedule applies.
| Model | OpenAI API | Azure OpenAI |
|---|---|---|
| o1, o3-mini, o4-mini | October 23 | November 19 |
| gpt-4o 2024-05-13 | October 23 | December 9 |
| gpt-4.1-nano | October 23 | April 14, 2027 |
| gpt-image-1 | October 23 | October 23 |
Azure answers every request with 410 Gone once a retirement date passes. Standard deployments get upgraded to a replacement region by region, but Provisioned deployments with reserved throughput receive no automatic upgrade and have to be migrated by hand.
Fine-tuned models have to be trained again
Swapping an ID will not revive a fine-tune. For ft-o4-mini and ft-gpt-4.1-nano the deprecations page lists only a base model as the replacement, which means retraining on the new base with the same data. That drew the loudest objections in the April announcement thread: one user wrote off the time invested in fine-tuning as lost, and another said the retirement was reason enough to move to Mistral or an open-source model.
If you run anything with o4-mini or gpt-3.5-turbo written into source or environment variables, today's first move is finding those IDs in your repositories and deployment configs.
grep -rnE "o4-mini|o3-mini|o1-pro|\"o1\"|gpt-3\.5-turbo|gpt-4-turbo|\"gpt-4\"|gpt-4o-2024-05-13|gpt-4\.1-nano|gpt-image-1" .
Where that turns something up, run about 50 of your usual inputs through GPT-6 Luna and through the officially named replacement, and compare output quality and cost per request side by side. If Luna holds up on your workload, the bill will be far smaller than it would be following the replacement table as written.