On 23 October 2026, OpenAI switches off most of its pre-2025 API lineup in one go: GPT-4, GPT-4 Turbo, o1, o1-pro, o3-mini, o4-mini, GPT-3.5 Turbo, GPT-4.1 nano and the first GPT-4o snapshot. Requests to those IDs fail after the date. The quick answer for each model is in the table below. The short version: if you run a large model (GPT-4, GPT-4 Turbo, o1), migrating cuts your bill by 70-90%. If you run a small one (o3-mini, o4-mini, GPT-3.5 Turbo), OpenAI’s named replacement costs two to four times more, and the cheaper path is a GPT-6 model OpenAI’s notice does not mention because it did not exist yet.
The replacement map, priced
OpenAI announced this shutdown on 22 April 2026 and named a substitute for each model. Every substitute is a GPT-5.6 model, because GPT-6 Sol and GPT-6 Luna did not ship until 22 September. The table prices each move on one workload: a support chatbot handling 5,000 conversations a day, 3,000 input and 400 output tokens each, 60% of input served from the prompt cache. All prices are OpenAI’s standard rates from its pricing page, checked 16 October 2026. All of these models use OpenAI tokenizers, so the token counts carry over unchanged.
| Retiring model | Monthly cost today | OpenAI’s named substitute | Monthly cost | Cheaper GPT-6 option | Monthly cost |
|---|---|---|---|---|---|
| gpt-4-0613 | $17,100 | gpt-5.6-sol | $2,028 (−88%) | gpt-6.1-sol | $987 (−94%) |
| gpt-4-turbo | $6,300 | gpt-5.6-sol | $2,028 (−68%) | gpt-6.1-sol | $987 (−84%) |
| gpt-4o-2024-05-13 | $3,150 | gpt-5.6-sol | $2,028 (−36%) | gpt-6.1-sol | $987 (−69%) |
| o1 | $8,325 | gpt-5.6-sol | $2,028 (−76%) | gpt-6.1-sol | $987 (−88%) |
| o3-mini | $611 | gpt-5.6-sol | $2,028 (+232%) | gpt-6.1-sol | $987 (+62%) |
| o4-mini | $536 | gpt-5.6-terra | $1,134 (+111%) | gpt-6.1-sol | $987 (+84%) |
| gpt-3.5-turbo-0125 | $315 | gpt-5.6-terra | $1,134 (+260%) | gpt-6-luna | $51 (−84%) |
| gpt-4.1-nano | $49 | gpt-5.6-luna | $113 (+133%) | gpt-6-luna | $51 (+4%) |
Two caveats on the GPT-5.6 Sol column. Its $4/$20 rate is a promotion that OpenAI says runs “at least through” 21 November 2026. At the $5/$30 list price, the same workload costs $2,835 a month, so a team that migrates to the named substitute now may face a second price change four weeks later. GPT-6.1 Sol has no such clock: it lists at $2/$10 with cached input at $0.10.
Run your own numbers in the API cost calculator, which has every model in this table preloaded. The retiring ones are flagged.
If you run GPT-4, GPT-4 Turbo, o1 or the 2024 GPT-4o snapshot
This is the easy case: every replacement is cheaper. The only real decision is GPT-5.6 Sol or GPT-6.1 Sol.
- The named substitute, gpt-5.6-sol, is the conservative choice: it is the model OpenAI’s own notice points to. Budget at its list price ($5/$30), not the promotion.
- gpt-6.1-sol costs about half as much on this workload, mostly because its cached input is $0.10 against $0.40. It is also the newer model: OpenAI shipped it on 29 September as its mid-tier recommendation, and lists it with a 1,050,000-token context window, the same as GPT-5.6 Sol.
- o1 users should check reasoning settings, not just the price. o1’s replacement path in OpenAI’s notice is gpt-5.6-sol (o1-pro maps to gpt-5.6-sol with
reasoning.mode: pro). Effort settings on the newer models can move cost per task as much as the rate card does, so set them explicitly rather than inheriting a default. - gpt-4-0613 users also get out of an 8,192-token context window. Prompts that were trimmed to fit can be restored, but every restored token is billed, so the saving above shrinks if you lengthen prompts.
One pricing rule is new: prompts over 272,000 input tokens on GPT-5.6 and GPT-6 models are billed at twice the input rate and 1.5 times the output rate for the whole request. Of the retiring models, only GPT-4.1 nano accepted prompts that long, at a flat rate, so this matters to long-context nano workloads and to anyone who plans to use the larger window.
If you run o3-mini, o4-mini, GPT-3.5 Turbo or GPT-4.1 nano
This is where the named substitutes hurt. OpenAI maps the small, cheap models to larger ones:
- o3-mini → gpt-5.6-sol. A small reasoning model replaced by a frontier one: roughly 3.3 times the monthly bill on this workload at the promotional rate, 4.6 times at list price.
- o4-mini → gpt-5.6-terra. $2/$12 against $1.10/$4.40: about double the bill.
- gpt-3.5-turbo → gpt-5.6-terra. About 3.6 times the bill.
- gpt-4.1-nano → gpt-5.6-luna. About 2.3 times the bill.
The alternative is GPT-6 Luna at $0.10 input and $0.50 output, with cached input at $0.01. On this workload it costs $51 a month against $315 for GPT-3.5 Turbo, and about the same as GPT-4.1 nano. It is the cheapest model in OpenAI’s GPT-6 lineup, so it is the obvious first test for anything that ran on GPT-3.5 Turbo or GPT-4.1 nano.
For o3-mini and o4-mini, there is no price-equivalent like-for-like. GPT-6 Luna is far cheaper than either but a much smaller model; GPT-6.1 Sol is the next rung up and costs 60-85% more than they did. The practical order is: run your evaluation set on GPT-6 Luna first, and move up to GPT-6.1 Sol only for the tasks it fails. If most requests pass on Luna, a split like that lands well below the named substitute’s bill.
Fine-tunes: the part with no substitute
The five fine-tune bases on the list — ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano-2025-04-14, ft-babbage-002 and ft-davinci-002 — stop serving on 23 October, and the fine-tuned models built on them stop with them. Re-training is harder than it was a year ago: OpenAI stopped new fine-tuning for organisations without recent fine-tuned inference on 2 July 2026, and existing customers lose the ability to create new jobs on 6 January 2027. If a fine-tune is load-bearing, start the re-training run this week on a base OpenAI still serves, or plan a prompting replacement and evaluate it before the date.
What else is retiring on the same timeline
- gpt-image-1 also shuts down on 23 October; OpenAI names gpt-image-2.5-sunburst or gpt-image-2.5-flare.
- gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini shut down on 14 January 2027, announced 14 October with 92 days’ notice, half the 184 days this batch received. If you are moving off GPT-4-era models anyway, it is cheaper to move those in the same project — the replacement map and the cache-rate rule that decides whether it costs you more are in migrating off the GPT-4 series.
- Agent Builder, the Evals platform and the reusable prompts API shut down on 30 November 2026.
Every date, with replacements and a subscribable calendar feed, is on the AI model retirement calendar.
A one-afternoon migration checklist
- Find every caller. Export usage by model from the OpenAI dashboard, and search code and configs for
gpt-4,gpt-3.5-turbo,o1,o3-mini,o4-miniandgpt-4.1-nano, including inside environment variables and model-routing tables. - Choose per workload, not per company. Large-model calls and small-model calls have opposite economics in this migration; price them separately.
- Evaluate on your own traffic. Run the same 100-200 real requests through the candidate models at the effort settings you will use, and compare outcome and cost per completed task.
- Pin a dated model ID where one exists. An alias moves when the vendor updates it, and a moved alias is how a migration turns into a regression.
- Ship before 21 October. That leaves two working days to roll back if a replacement misbehaves before the old IDs stop answering.
What changed
- 16 Oct 2026: Guide published, with prices checked against OpenAI’s pricing and deprecation pages the same day.
Frequently asked questions
Which OpenAI models shut down on 23 October 2026?
gpt-4-0613 (with the gpt-4 alias and its completions variants), gpt-4-turbo, gpt-4-1106-preview, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini, gpt-3.5-turbo-0125 (with the gpt-3.5-turbo alias), gpt-4.1-nano and gpt-image-1, plus the fine-tune bases ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano, ft-babbage-002 and ft-davinci-002. OpenAI announced all of them on 22 April 2026, 184 days ahead. After the date, requests to these IDs fail.
Is gpt-4o being shut down on 23 October?
No. Only the gpt-4o-2024-05-13 snapshot goes on 23 October. The gpt-4o alias, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini were deprecated separately on 14 October 2026 and shut down on 14 January 2027, 92 days after that notice.
Should I migrate to the replacement OpenAI names, or to GPT-6?
OpenAI's substitute list names GPT-5.6 models because it was written on 22 April, five months before GPT-6 Sol and GPT-6 Luna shipped on 22 September. On a 5,000-conversation-a-day chatbot, GPT-6.1 Sol costs about half of the named GPT-5.6 Sol substitute, and GPT-5.6 Sol's $4/$20 rate is a promotion OpenAI only guarantees through 21 November 2026 (list price $5/$30). Test both on your own evaluations, but price the GPT-6 option before accepting the default.
Why does my o4-mini bill go up after migrating?
Because the named substitute is a bigger, dearer model. o4-mini costs $1.10 per million input tokens and $4.40 per million output; OpenAI's substitute, gpt-5.6-terra, costs $2 and $12. On the chatbot workload priced on this page that is roughly double the monthly bill. GPT-6 Luna at $0.10/$0.50 is far cheaper but a much smaller model, so it only works if it passes your quality bar.
What happens to fine-tuned GPT-3.5 Turbo and GPT-4 models?
They stop working on 23 October together with their base models (ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano-2025-04-14, ft-babbage-002, ft-davinci-002). OpenAI has also narrowed who can create new fine-tuning jobs: since 2 July 2026 only organisations that ran fine-tuned inference in the previous 60 days can, and on 6 January 2027 existing customers lose that too. If you depend on a fine-tune, re-train it now or plan to replace it with prompting.