In this guide · 7 sections
Short answer
OpenAI's fine-tuning platform closed to new organisations on 7 May 2026, and from 6 January 2027 no customer can create new fine-tuning jobs. Existing fine-tuned models work until their base model shuts down, which for GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini is 14 January 2027.
What OpenAI announced, and when
OpenAI added a dated note to its 2024 post on fine-tuning API improvements: "Update on May 8, 2026: OpenAI is winding down the fine-tuning platform." The note says the platform is no longer open to new users, existing users can create training jobs "for the coming months", and fine-tuned models stay available for inference until their base models are deprecated.
The exact timeline is on OpenAI's deprecations page, under "Update to OpenAI's self-serve fine-tuning". It says developers were notified on 7 May 2026.
| Date | What changes | Who it affects |
|---|---|---|
| 7 May 2026 | Creating fine-tuning jobs is not available to organisations that have not previously run fine-tuning | New users |
| 2 July 2026 | Job creation removed for organisations that have not run inference on a fine-tuned model in the past 60 days | Inactive past users |
| 6 January 2027 | Active existing customers can no longer create new fine-tuning jobs | Everyone |
| Base model shutdown | Inference on a fine-tuned model is disabled when its base model is deprecated | Anyone still using a fine-tuned model |
When your fine-tuned model actually stops working
Training is one deadline; inference is the one that breaks production. Because fine-tuned models follow their base model, the base model's shutdown date is your real cut-off. The deprecations page lists several that matter to small businesses:
- 23 October 2026: fine-tuned gpt-3.5-turbo, gpt-4, gpt-4.1-nano (2025-04-14), o4-mini (2025-04-16), babbage-002 and davinci-002 models are removed, per the 22 April 2026 legacy snapshot notice.
- 14 January 2027: GPT-4o (2024-08-06 and 2024-11-20), GPT-4o mini (2024-07-18), GPT-4.1 and GPT-4.1 mini (2025-04-14) are removed, and the page says "Fine-tuned versions of these models follow the same shutdown schedule" (notice of 14 October 2026).
OpenAI names a recommended replacement base model for each shutdown, but those are base models, not your fine-tune. Whatever behaviour you trained in has to be rebuilt another way.
Why most small businesses do not need fine-tuning
Fine-tuning changes how a model behaves by training it on example conversations. It does not reliably teach a model facts, and it has to be redone when prices or policies change. Most SME requests fall into three groups, and each has a simpler fix that does not depend on a training platform.
| What you wanted | Fine-tuning approach | Better approach in 2026 |
|---|---|---|
| Answers from our prices, FAQs and policies | Train on Q&A pairs | Retrieval (RAG): the model looks up your current documents before answering |
| Our tone and wording | Train on past replies | A clear system prompt with 5 to 10 example replies |
| A fixed output format (quotes, JSON for another system) | Train on formatted examples | Structured outputs: the API enforces a JSON schema you define |
| Classifying emails or tickets | Train a classifier | Prompt with labelled examples, checked against a test set |
Our fine-tuning vs RAG guide covers the trade-offs in more depth, and RAG explained shows how a chat-with-your-documents assistant works. OpenAI documents structured outputs as a way to make the model follow a JSON schema you supply, which covers most "always answer in this format" cases.
Migration plan for existing fine-tuning users
If you already run a fine-tuned model, work through these steps before the base model's shutdown date, not on it.
List every fine-tuned model ID in use
Search code, environment variables and no-code tools for "ft:". Note the base model in each ID and its shutdown date on the deprecations page.
Export your training and evaluation data
Keep the JSONL training files and, more importantly, a test set of real inputs with the outputs you expect. You will need them whichever route you take.
Rebuild with prompts and examples on a current model
Put the behaviour into a system prompt with examples, and move facts into retrieval. Run your test set and compare against the fine-tune.
Use structured outputs for format-sensitive jobs
If the fine-tune existed to produce valid JSON or a strict template, define a schema and let the API enforce it.
Decide whether you still need a training run
If results are close enough, stop there. If not, and you are still an active customer, you can train on a newer supported base before 6 January 2027, knowing that model will also retire one day.
Switch, monitor, then delete the old model
Roll out the replacement, compare outputs for a week, then remove the old model ID from your code.
If you truly need custom training: other platforms
A small number of cases do benefit from training, for example a high-volume classifier where a smaller tuned model is cheaper than a large general one. OpenAI's own platform will not accept new jobs after 6 January 2027, but other providers still document fine-tuning as of October 2026:
- Microsoft Foundry (Azure OpenAI): Microsoft's supervised fine-tuning guide covers training jobs on supported models and regions. Check the supported model list before planning, as it changes.
- Google Vertex AI: Google's introduction to tuning documents supervised tuning for Gemini models.
- Open-weight models: you can fine-tune and host an open model yourself, which adds hosting, security and update work.
Each route ties you to a provider's model lifecycle again. Our Azure OpenAI vs OpenAI API comparison covers the contract and data-residency differences if you are weighing Microsoft's platform.
Do it yourself or get help?
If your fine-tune was a weekend experiment, the simplest move is to stop using it and rewrite the prompt. If it sits inside a customer-facing chatbot, a quoting tool or an integration, plan the replacement properly and test it against real inputs before the shutdown date.
Before the base model retires
IT Sorted is a Singapore-based team of developers with over 10 years of experience. Our AI Website Assistant (S$790) builds a chat assistant that answers from your own website, FAQs and documents using retrieval, with guard-rails and lead capture. For a replacement inside a custom workflow or integration, Custom Tool or Dashboard (from S$1,990) is scoped in writing with a fixed price before we start. AI usage fees are billed to you by the provider.
Frequently asked questions
Only if your organisation is an active existing fine-tuning customer. OpenAI's deprecations page says organisations that had never fine-tuned lost access on 7 May 2026, those without fine-tuned inference in the previous 60 days lost job creation on 2 July 2026, and all customers lose the ability to create new jobs on 6 January 2027.
Yes, until its base model is deprecated. OpenAI says inference on fine-tuned models is disabled only when the underlying base model shuts down. For example, fine-tuned GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini models follow those base models' removal on 14 January 2027, per the deprecations page.
Retrieval-augmented generation, usually called RAG. The assistant searches your current documents, FAQs and price lists and answers from them, so updates take effect as soon as you change the source. Combine it with a clear system prompt for tone. This avoids retraining and is not tied to a training platform's lifecycle.
For most formatting needs, yes. OpenAI's structured outputs feature lets you supply a JSON schema and have the model's reply follow it, which covers quotes, extraction into fields and data passed to another system. Fine-tuning was often used for this before; a schema is simpler and works with current models.
As of October 2026, Microsoft documents supervised fine-tuning in Microsoft Foundry for supported models and regions, and Google documents tuning for Gemini models in Vertex AI. Supported models change over time, so check each provider's current model list and lifecycle before you invest in a training run.