OpenAI is winding down its fine-tuning platform. Existing customers can create new fine-tuning jobs until 6 January 2027, and the GPT-4.1 and GPT-4o models most teams fine-tuned shut down, fine-tunes included, on 14 January 2027. The practical alternative is to retrain the same task on an open model, serve it behind an OpenAI-compatible API, and evaluate it before moving traffic.
This guide covers what is ending and when, what happens to the models you already have, the options open to you, a four-step migration, and how to work out what the switch costs. Every date and price below was read from the vendor's own documentation on 28 October 2026.
What is OpenAI ending, and when?
OpenAI's deprecations page sets out the fine-tuning wind-down in three stages, followed by the shutdown of the base models the remaining fine-tunes run on. Its supervised fine-tuning guide now opens with the line "OpenAI is winding down the fine-tuning platform."
| Date | What changes |
|---|---|
| 7 May 2026 | Organisations that never fine-tuned can no longer start |
| 2 July 2026 | Organisations with no fine-tuned inference in 60 days lose job creation |
| 23 October 2026 | GPT-4.1 nano, o4-mini and their fine-tunes shut down |
| 6 January 2027 | Remaining customers can no longer create fine-tuning jobs |
| 14 January 2027 | GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down |
The last row is the one most teams miss. OpenAI's rule is that inference on a fine-tuned model continues until its base model is deprecated, and on 14 October 2026 OpenAI announced that those four base models leave the API on 14 January 2027, with "fine-tuned versions of these models" following "the same shutdown schedule". In practice that gives most fine-tuned GPT models eight days of life after training closes.
What happens to your existing fine-tuned models?
A fine-tuned OpenAI model keeps answering API calls until its base model shuts down, and then it stops. For a fine-tune of GPT-4o, GPT-4o mini, GPT-4.1 or GPT-4.1 mini, that date is 14 January 2027. Fine-tunes of GPT-4.1 nano and o4-mini were already shut down on 23 October 2026.
OpenAI names GPT-6 Luna as the recommended replacement for all four models. GPT-6 Luna is a general model, so the behaviour your fine-tune learned from your examples does not carry over. You get it back in one of two ways: by putting the instructions and examples into every prompt, or by training a new model on the same data somewhere else.
What you keep is your training data, your evaluation set and your production logs. The fine-tuned model itself lives on OpenAI's API, so plan the migration around the data. For the wider pattern of forced migrations, see what happens when OpenAI deprecates the model your app depends on.
What are the alternatives to OpenAI fine-tuning?
There are three routes. They differ in who runs the training, who serves the model and how long the base model is guaranteed to stay available.
| Route | Who trains | Who serves | Main risk |
|---|---|---|---|
| Another closed provider | The provider | The provider | Base model retires on their schedule |
| Managed training on open models | A training platform | The platform or your cloud | Choosing a capable partner |
| Self-hosting an open model | Your team | Your team | Engineering and operations load |
The right route depends on how much of the training and serving work you want to own. Most teams that fine-tuned on OpenAI did so because they wanted to skip that work, which makes the second route the closest match.
Other closed providers
Google still offers supervised tuning for Gemini, through Gemini Enterprise Agent Platform. As of 28 October 2026, Google's tuning documentation lists Gemini 3.5 Flash, Gemini 3.5 Flash-Lite, Gemini 3.1 Flash-Lite and three Gemini 2.5 models. The developer Gemini API and AI Studio dropped tuning with the deprecation of Gemini 1.5 Flash-001 in May 2025, and Google says it has no immediate plans to bring it back.
Mistral's fine-tuning API now sits under "Deprecated features" in its documentation, labelled legacy and "no longer actively supported" as of October 2026.
A closed provider is the shortest move if you already run on that cloud. The trade is the one you are living through now: the base model has a retirement date set by someone else. Google lists Gemini 3.5 Flash for retirement on "May 19, 2027 or later" as of October 2026.
Managed training on open models
A managed platform trains an open model on your examples and serves it for you, usually through an OpenAI-compatible API. An open base model has no vendor-set retirement date, so the trained version can run for as long as you need it. You hand over training data and an evaluation set, agree a quality bar, and get an endpoint.
This route keeps the convenience of OpenAI fine-tuning and removes the retirement risk. The questions to ask a provider are where the model can run (their infrastructure, your cloud, or on-premise), how a new version gets approved, and how pricing is set.
Self-hosting an open model
Your team picks a base model, runs the training and serves the result on its own infrastructure. It gives you the most control and the most work: training pipelines, serving, scaling, monitoring and upgrades all become yours. Our 90-day plan for moving to a fine-tuned local model walks through this route end to end.
How do you migrate off OpenAI fine-tuning?
The migration has four steps, and the order matters: gather the data first, and move traffic last.
1. Export your training data
Collect the JSONL files you used for each fine-tuning job, from wherever your team stored them. Add a recent sample of production prompts and the fine-tuned model's responses from your own logs, because they show the task as it runs today, including the inputs your original examples never covered.
Set aside a held-out evaluation set before you do anything else, and never train on it. It is the only fair way to compare the old model with the new one.
2. Re-train on an open model
Train an open model on the same examples. The OpenAI chat format (a list of system, user and assistant messages per example) converts directly to the format most training tools accept, so the data work is mostly cleaning and deduplication.
Start from the smallest model that holds your quality bar on the evaluation set. A model trained for one task can be much smaller than a general model and still match it on that task, which is where most of the running-cost saving comes from.
3. Keep the OpenAI-compatible API shape
Serve the new model behind an OpenAI-compatible API, so your application code changes in two places: the base URL and the model name.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://your-endpoint.example.com/v1",
api_key=os.environ["ENDPOINT_API_KEY"],
)
response = client.chat.completions.create(
model="your-trained-model",
messages=[{"role": "user", "content": ticket_text}],
)Your prompts, retries, structured-output parsing and logging stay as they are. That keeps the migration risk inside the model, where your evaluation set can measure it.
4. Evaluate before you switch traffic
Run the old fine-tune and the new model on the held-out set and compare them on the measures your product depends on: exact-match accuracy, schema validity, refusal rate, latency. Then shadow live traffic, send a small share to the new model, and widen the share only while it holds the bar. Our guides on evaluating a fine-tuned model and A/B testing a fine-tuned model in production cover both stages.
Finish this step before 14 January 2027, while the old fine-tune is still there to compare against.
What does it cost to migrate off OpenAI fine-tuning?
Cost per task has two parts: the price per token and the number of tokens each task uses. A fine-tune keeps the second number small, because the instructions and examples are trained into the model. Remove the fine-tune and they go back into every prompt.
The worked example below is a classification task of 600 input tokens and 150 output tokens, run 1 million times a month, at OpenAI's standard list prices as of October 2026 from its pricing page. The second and third rows add 3,000 tokens of instructions and examples to each prompt to stand in for the lost fine-tune.
| Setup | Input tokens per task | Monthly cost |
|---|---|---|
| Fine-tuned GPT-4.1 mini | 600 | $960 |
| GPT-6 Luna with instructions in the prompt | 3,600 | $435 |
| GPT-6 Luna with the instructions cached | 3,600 | $165 plus cache writes |
The rates behind the table: fine-tuned GPT-4.1 mini costs $0.80 per million input tokens and $3.20 per million output tokens; GPT-6 Luna costs $0.10 and $0.50, with cached input at $0.01. GPT-6 Luna comes out cheaper even carrying the longer prompt, so for many tasks the cost question settles quickly. The open question is quality. The cheapest row only counts if it holds your bar, which is why step 4 comes before any bill comparison.
A trained model keeps the 600-token shape after the fine-tune is gone, and its price per token decides the rest. Run the same arithmetic for every candidate, with your own token counts from your logs. Our cost-per-task calculator does it for you and shows how much of the difference comes from the price per token and how much from the tokens each task sends.
Where does Ertas fit?
Ertas runs an adaptive model endpoint: a model trained on your task, served through an OpenAI-compatible API, which you reach by changing the base URL and model name as in step 3. Ertas manages it, or it runs in your own cloud or on-premise. Each new version is checked against your evaluation set, and your team approves it before it takes traffic. The endpoint is currently invitation only, and we are developing a more streamlined version for wider access.
The work starts with a free audit of where your AI spend goes and which tasks repeat. The build is a fixed fee scoped after that audit, and running costs are set per token, as a flat monthly seat or on dedicated hardware, as listed on our pricing page. If you would rather run the training yourself, Ertas Studio trains models on managed infrastructure and lets you download the finished files.
Moving a fine-tuned GPT model before 14 January 2027? See how the adaptive endpoint works and book a call from there.
Frequently Asked Questions
Is fine-tuning still used?
Yes. Teams fine-tune open models every day to make a model reliable on one task, and Google still offers supervised tuning for Gemini through Gemini Enterprise Agent Platform as of October 2026. What is shrinking is self-serve fine-tuning of closed models: OpenAI closes new jobs on 6 January 2027, and Mistral lists its fine-tuning API as deprecated.
How much does LLM fine-tuning cost?
On OpenAI, as of October 2026, supervised fine-tuning is billed at $5 per million training tokens for GPT-4.1 mini and $25 for GPT-4.1 and GPT-4o, plus per-token inference on the result. Self-serve training on open models in Ertas Studio starts free, with paid plans from $10 a month as of October 2026. A managed build is usually a fixed fee set by the size of the task.
When does OpenAI stop fine-tuning?
Existing customers can create new fine-tuning jobs until 6 January 2027. New organisations lost access on 7 May 2026. Fine-tuned GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini models shut down with their base models on 14 January 2027.
What should I replace a fine-tuned GPT-4.1 mini model with?
Retrain the same task on an open model and serve it behind an OpenAI-compatible API, or move to GPT-6 Luna, OpenAI's recommended replacement, with your instructions in the prompt. Compare both on a held-out evaluation set before 14 January 2027, while the original fine-tune is still available as a baseline.
Sources, read 28 October 2026: OpenAI deprecations, supervised fine-tuning guide and API pricing; Google Gemini API fine-tuning and Gemini Enterprise Agent Platform supervised tuning; Mistral fine-tuning API (legacy). For the comparison of the two products, see Ertas vs OpenAI fine-tuning.