Fine-Tuned Model Migration
OpenAI stops new fine-tuning jobs for all customers on 6 January 2027, and fine-tuned GPT-4o and GPT-4.1 models shut down on 14 January 2027. We test OpenAI's replacement on your own examples first. If it falls short, we rebuild your fine-tune on a model you control, before the date.
Bring your fine-tuned model ID. We'll tell you what testing the replacement involves.
If the replacement model is good enough, we'll tell you, and the work stops there.
We rebuild and hand over the model. Running it in production is a separate engagement.
What happens to fine-tuned OpenAI models?
OpenAI is winding down self-serve fine-tuning. Your fine-tuned model keeps working until its base model is retired. For fine-tunes of GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini, that's 14 January 2027. OpenAI recommends gpt-6-luna as the replacement, but that is a general model: the behaviour your fine-tune learned from your examples doesn't come with it. OpenAI documents no way to download fine-tuned weights. What you can take with you is your training data, your evaluation examples and your production logs, and that's what a migration is built from. It's one of three services in our AI model migration and modernization work. If your model isn't fine-tuned, see AI model migration.
What's ending, and when
| Date | What happens |
|---|---|
| 7 May 2026 | No new fine-tuning for organisations that had never fine-tuned |
| 2 Jul 2026 | No new jobs for organisations without fine-tuned inference in the previous 60 days |
| 23 Oct 2026 | Fine-tuned o4-mini (ft-o4-mini-2025-04-16) shut down |
| 31 Oct 2026 | OpenAI Evals: existing evals become read-only |
| 30 Nov 2026 | OpenAI Evals dashboard and API shut down |
| 6 Jan 2027 | No new fine-tuning jobs for any customer |
| 14 Jan 2027 | GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down, with their fine-tunes |
Source: OpenAI deprecations page, checked 8 November 2026. If you kept your evaluation sets in OpenAI Evals, export them now: they're needed for step 2 below.
OpenAI also says some customers may be able to buy dedicated capacity to keep using a model after its shutdown date, through its sales team. If that buys you the time you need, it's worth asking.
Your four options
| Option | What it means | Good when | Watch out for |
|---|---|---|---|
| 1. Switch to the recommended replacement | Move to gpt-6-luna with your best prompt and examples | It matches your fine-tune on your evaluation set | Behaviour that was trained in may not come back from a prompt |
| 2. Ask OpenAI for dedicated capacity | Keep the current fine-tune running past the date | You need more time, and the cost works | It delays the decision rather than making it |
| 3. Rebuild on a managed platform that still offers tuning | Re-train on e.g. Gemini (Google), Llama or Nova (AWS Bedrock), or a GPT model on Microsoft Foundry, which still documents GPT-4.1 fine-tuning (page updated October 2026; check it's open to you) | You want managed hosting and that provider's models fit | The base model has a retirement date set by someone else again |
| 4. Rebuild on an open-weight model you control | Re-train an open model in your own cloud or on-premises | You want to own the next retirement date | Your team runs the model; licences vary |
We test option 1 first, because if it works it's the cheapest. The fit memo then recommends one of the others if it doesn't.
Migration checklist: start this week
- 01Download what you own. Training and validation files, the hyperparameters of your fine-tuning jobs, prompts and system messages, and a sample of production inputs with your fine-tune's outputs. Check that every file opens.
- 02Build an evaluation set. A few hundred real examples your fine-tune handles well and a few it gets wrong, with the expected answer for each. Keep it out of any training. If it lived in OpenAI Evals, export it before 30 November 2026.
- 03Score the replacement. Run your fine-tune and
gpt-6-luna(with the best prompt you can write) on the same set. Compare what your product depends on: exact-match accuracy, schema validity, refusal rate, latency, cost per call. - 04Decide. If the replacement holds up, switch with the evidence. If not, choose option 3 or 4.
- 05Rebuild, then shadow. Re-train on the chosen base model, keep the same OpenAI-compatible API shape so your application code barely changes, score it on the same set, then send a copy of live traffic to it without showing users the answers before you switch.
Steps 1 and 2 need nobody but your team. We can run all five, or pick up at step 3.
What you end up with
- A decision made on your data, not on a launch blog post: old and new models scored side by side on your own examples.
- The same behaviour after the deadline, either from the replacement or from a rebuilt model.
- Little application change, because the API shape stays the same.
- No forced migration next time if you choose an open-weight base: no vendor sets its retirement date, and you keep the pipeline to re-run.
- An evaluation set you keep, which outlives OpenAI Evals and works on any future model.
How a migration runs
| Stage | What happens | You get |
|---|---|---|
| Migration Sprint | Checklist steps 1–4: collect, build the evaluation set, score your fine-tune against the replacement, recommend | Migration memo with scores and a deadline plan. Gate 1: you choose the option |
| Rebuild (only if needed) | Data, training on the chosen base model, scoring on the same set, regression checks, cost and latency | Evaluation report. Gate 2: you accept it |
| Hand over | Package, OpenAI-compatible endpoint config, runbook, shadow-traffic plan; your team re-runs the pipeline once | Deployment package and runbook |
If the Migration Sprint shows the replacement is good enough, the engagement ends there.
How you buy it: free fit call → Migration Sprint → Rebuild (only if needed). The Migration Sprint and the Rebuild are each priced per scope.
Working back from 14 January 2027: leave time for shadow traffic before the switch.
What you receive
- 01Migration memo: scores for your fine-tune and the replacement, the recommended option, licence notes for any new base model, and a plan to the deadline.
- 02Evaluation set and baseline, yours to keep.
- 03Training dataset and pipeline (if rebuilding).
- 04Rebuilt model: weights or adapter, configurations and run log (if rebuilding).
- 05Evaluation report: old vs new on task score, schema validity, refusals, latency and cost per call.
- 06Deployment package: model files, an OpenAI-compatible serving config, a runbook and a handover session.
You own the trained model, the dataset and the evaluation set once the work is paid for; we keep our own reusable tooling, and the base model's licence still applies. Need it connected to your product? We can scope that integration into the Build.
What we need from you
Your fine-tuned model ID and its base model · your training and validation files (or access to download them) · prompts and a sample of production traffic · someone who knows what a correct answer looks like · where the new model must run · your real deadline.
Not included
Hosting the rebuilt model (your team runs it, with the runbook we hand over) · negotiating dedicated capacity with OpenAI (you contact their sales team) · rewriting your application beyond pointing it at a new endpoint · migrating non-fine-tuned API usage to new models.
How to judge us
The Migration Sprint ends with a recommendation even if it's “switch to GPT-6 Luna and stop”.
| Question | Crescent AI | Migration guide from a tooling vendor | Doing it in-house |
|---|---|---|---|
| Tests the replacement before rebuilding | Yes, it's step 3 | Sometimes | If you have the time |
| Evaluation set built from your traffic | Yes | You build it | You build it |
| Recommends options that don't involve us | Yes (replacement, dedicated capacity) | Usually points to its platform | n/a |
| Licence review of the new base model | Yes | Rarely | Your legal team |
| Best when | The deadline is close and you want one team to decide and deliver | You have engineers free to follow a guide | You have engineers free and an evaluation set already |
Who leads this work
Yash Amin, founder of Crescent AI. About Crescent AI
Common questions
Can I export my fine-tuned model's weights from OpenAI?
OpenAI documents no way to do this. Plan around what you can download: training data, evaluation examples and production logs.
What should I replace a fine-tuned GPT-4.1 mini model with?
OpenAI recommends gpt-6-luna. Test it with your best prompt on your own examples first. If it falls short, rebuild on an open-weight or managed model.
Does my fine-tuned model stop working on 6 January 2027?
No. On 6 January you can no longer create new fine-tuning jobs. Inference continues until the base model shuts down: 14 January 2027 for the GPT-4o and GPT-4.1 families.
Can GPT-5 or GPT-6 models be fine-tuned instead?
OpenAI is winding down self-serve fine-tuning for all customers, so plan on no. Check with OpenAI if you have an enterprise agreement.
What if we use Azure OpenAI rather than OpenAI's API?
Microsoft runs its own schedule: its Foundry documentation (updated October 2026) still covers fine-tuning GPT-4.1, though some models are limited to existing customers. Check its documentation and your agreement. The same checklist applies if your model is being retired there.
Other vendors are retiring models too. Does this apply?
Yes. The same five steps work for any retired fine-tuned model or tuning service.
Not using OpenAI?
Any provider can retire a tuned model or its tuning service. The process is the same: download what you own, build an evaluation set, test the provider's replacement, and rebuild only if it falls short. To fine-tune a new model from scratch rather than migrate one, see AI model fine-tuning.
Bring us your fine-tuned model ID and its shutdown date.
We'll tell you whether the replacement is likely to hold up, and what testing it on your examples involves.
Mutual NDA before you share data · Replacement tested before any rebuild · Fixed-scope Migration Sprint
Part of AI Model Migration & Modernization · AI Model Migration · AI Model Fine-Tuning · AI Evaluation