OpenAI fine-tuning deprecated: what to use instead in 2026

The date most teams have noted for OpenAI's fine-tuning wind-down is 6 January 2027. For anyone running a fine-tuned o4-mini or GPT-4.1 nano model, the date that matters is 23 October 2026, when those models stop answering. With OpenAI fine-tuning deprecated in stages, inference on a fine-tune lasts only as long as its base model. This post gives the full timeline, what each alternative costs on published prices, and how to choose per use case.

If your product depends on a fine-tuned OpenAI model, you have two deadlines, not one: the day you can no longer retrain it, and the day its base model is retired. The second is the one that takes a feature offline.

OpenAI fine-tuning deprecated: the dates, by who is affected

OpenAI notified self-serve fine-tuning customers on 7 May 2026. The deprecations page sets out the stages as of September 2026:

Date	What changes	Who is affected
7 May 2026	Creating fine-tuning jobs or training is unavailable	Organisations that had never run fine-tuning
2 July 2026	Creating fine-tuning jobs is unavailable	Organisations with no inference on a fine-tuned model in the past 60 days
23 October 2026	Shutdown of ft-o4-mini-2025-04-16, ft-gpt-4.1-nano-2025-04-14 and older fine-tunes of gpt-3.5-turbo, gpt-4, babbage-002 and davinci-002	Anyone serving those fine-tunes
6 January 2027	No new fine-tuning jobs for anyone	Active existing customers
Base model retirement	Inference on fine-tuned models disabled	Each fine-tune, on its own base model's date

The model optimization guide says the same in plainer words: the platform is no longer accessible to new users, existing users can create training jobs for the coming months, and all fine-tuned models remain available for inference until their base models are deprecated.

Will my fine-tuned model stop working?

Yes, eventually, and for some models next month. A fine-tuned model stops working when OpenAI retires its base model. The 23 October 2026 shutdown list includes o4-mini-2025-04-16 and gpt-4.1-nano-2025-04-14 alongside their fine-tuned versions, with gpt-5.6-terra and gpt-5.6-luna as the recommended replacements. Fine-tunes on gpt-4.1, gpt-4.1-mini and gpt-4o-2024-08-06 have no shutdown date on the page today.

That second group is not safe indefinitely, but it does have a floor. OpenAI's stated policy is at least six months' notice before retiring a generally available model. The October shutdowns were announced on 22 April, which fits that pattern.

Can I still fine-tune GPT models?

Only if your organisation is an active existing customer, and only until 6 January 2027. The guide lists supervised fine-tuning and direct preference optimisation on the gpt-4.1 family, vision fine-tuning on gpt-4o-2024-08-06, and reinforcement fine-tuning on o4-mini-2025-04-16 only. That last base model is on the 23 October shutdown list, so reinforcement fine-tuning has, in practice, a shorter runway than the rest.

What a fine-tune was buying you

Before picking a replacement, name what the fine-tune did. OpenAI's own guide lists four benefits over prompting alone: more examples than fit in one context window, shorter prompts that save tokens and latency at scale, training on sensitive data without sending it in every request, and making a smaller, cheaper model good at one task. Each has a different replacement.

What the fine-tune did	Replacement to test first
Consistent format or classification	A stronger base model with few-shot examples in a cached prefix
Domain knowledge	Retrieval over your documents
Shorter prompts to cut cost	Prompt caching, then a cheaper base model
A small model doing one narrow task	Distillation into a model you host

OpenAI fine-tuning alternatives, costed on published prices

The replacements OpenAI names are cheaper on list price in one case and more expensive in the other. All rates below are standard per-million-token prices from the OpenAI pricing page on 21 September 2026. The workloads are our assumptions, stated as such.

From fine-tuned o4-mini to GPT-5.6 Terra

A fine-tuned o4-mini bills at $4.00 input and $16.00 output per million tokens. Terra lists at $2.00 and $12.00. Assume 100,000 calls a month at 2,000 input and 500 output tokens: the fine-tune costs 200M x $4.00 plus 50M x $16.00, which is $1,600. Terra costs 200M x $2.00 plus 50M x $12.00, which is $1,000. That comparison assumes the same token counts; reasoning models vary in how many output tokens they spend, so measure it on your own traffic before trusting the saving.

From fine-tuned GPT-4.1 nano to GPT-5.6 Luna

This one goes the other way. A fine-tuned gpt-4.1-nano bills at $0.20 input and $0.80 output; Luna lists at $0.20 input and $1.20 output. Assume 1,000,000 calls a month at 500 input and 100 output tokens. The fine-tune costs $100 plus $80: $180. Luna with the same short prompt costs $100 plus $120: $220.

The catch is that the fine-tune needed no examples, and Luna may. Add a 2,000-token few-shot prefix, read from cache on 99% of calls and written on 1%, and Luna costs $264.60. Make the prefix 10,000 tokens and it costs $443.00, nearly two and a half times the fine-tune.

Setup (1M calls a month)	Arithmetic	Monthly
ft-gpt-4.1-nano, no prefix	500M x $0.20 + 100M x $0.80	$180.00
Luna, no prefix	500M x $0.20 + 100M x $1.20	$220.00
Luna, 2,000-token cached prefix	$220 + 1.98B x $0.02 + 20M x $0.25	$264.60
Luna, 10,000-token cached prefix	$220 + 9.9B x $0.02 + 100M x $0.25	$443.00

The fine-tuning premium you stop paying

One line in the comparison is easy to miss. Fine-tuned models have always billed above their base model: on the same pricing page, a fine-tuned gpt-4.1-mini costs $0.80 input and $3.20 output per million tokens, against $0.40 and $1.60 for the base gpt-4.1-mini. That is exactly double on both lines. A fine-tuned gpt-4.1 bills $3.00 and $12.00 against $2.00 and $8.00 for the base. When you move off a fine-tune you lose the training, but you also stop paying that premium, and the saving is what funds the longer prompt that replaces it. Run the comparison with the fine-tuned rate, not the base rate, or you will understate what you are saving.

Replace fine-tuning with prompt caching, carefully

Prompt caching is the lever that makes a long instruction block affordable. For GPT-5.6 models, the prompt caching guide bills cache reads at 0.1x the uncached input rate and writes at 1.25x, with a 1,024-token minimum prefix and a lifetime of at least 30 minutes after the last use. It turns examples you used to train into examples you send, at a tenth of the price.

The table above shows the limit. At high volume with tiny prompts, a tenth of the price of many tokens is still more than nothing, and a narrow fine-tune on a cheap model was hard to beat on cost. Caching is the right replacement when the fine-tune existed for quality or format; it is the wrong one when the fine-tune existed to make a very high-volume call cheap.

Fine-tuning vs RAG, and when to distil

If the fine-tune was teaching the model facts about your products, policies or catalogue, it was the wrong tool already. Retrieval-augmented generation, where you fetch relevant documents at request time and put them in the prompt, keeps knowledge current without retraining and lets you show where an answer came from. Fine-tuning was always better at behaviour than at knowledge; the deprecation just removes the option to pretend otherwise.

Distillation is the answer for the case caching loses: a narrow, high-volume task where a small model was the point. You generate labelled outputs from a strong model, check them, and train a smaller open-weight model you host yourself. You own the weights, and no base-model retirement can switch it off. The cost is that you now run inference infrastructure, with the capacity planning, monitoring and upgrades that come with it. We set out when that trade is worth making in custom machine learning solutions: when they beat an API, and our analysis of open-weight models against frontier spend shows where the savings tend to come from. Check your model provider's terms on using outputs as training data before you start.

The strongest argument for staying put is real: if your fine-tune sits on gpt-4.1 or gpt-4.1-mini, it has no announced shutdown and at least six months' warning when one comes. Retraining it now, before 6 January, buys time. It does not buy a future, because after that date you cannot retrain it again when your data or requirements change.

What to do before 23 October

Every replacement above needs the same thing first: an evaluation set built from real traffic that scores the fine-tune today, so you can tell whether the replacement is as good. Build it while the fine-tune still answers. If you use a model as the grader, correct for the biases we covered in LLM as a judge evaluation.

1. List every fine-tuned model ID in production and its base model. Anything on o4-mini or gpt-4.1-nano has until 23 October.
2. Capture a few thousand real inputs and the fine-tune's outputs now, as your baseline and your future distillation data.
3. Test the recommended replacement with a cached few-shot prefix, and cost it at your volume with the arithmetic above.
4. Where quality holds and cost does not, plan distillation; where knowledge was the goal, move to retrieval.

As an OpenAI Select Partner that also builds on Claude, Gemini and custom models, we are not tied to one answer, and our AI engineering services team starts every one of these migrations with that baseline. The decision for this week is simpler: find out which of your fine-tunes sit on a base model that retires on 23 October.