What OpenAI actually changed
OpenAI's deprecations page lists "self-serve fine-tuning job creation" as a restricted feature with a shutdown date of 6 January 2027, announced on 7 May 2026. The timeline underneath has three steps. On 7 May 2026, organizations that had never run fine-tuning lost the ability to create jobs. On 2 July 2026, so did organizations that had not run inference on a fine-tuned model in the past 60 days. On 6 January 2027, "active existing customers will no longer be able to create new fine-tuning jobs"1.
The same page states what does not change on that date: inference on fine-tuned models continues until the underlying base model is deprecated, and fine-tuned versions of listed models follow the same shutdown schedule as their base1. So two clocks run separately. The job clock ends 6 January 2027. The serving clock ends on the base model's shutdown date.
From today, 18 November 2026, the first clock has 49 days left. For tunes built on the GPT-4o and GPT-4.1 snapshots, the second one ends 14 January 2027, 57 days out, with gpt-6-luna as OpenAI's recommended replacement1. That replacement is listed as not supporting fine-tuning2. The base-retirement angle is covered in your fine-tuned model's base is retiring; this piece is about the other decision: what you do with the ability to tune at all.
What happens to existing tuned models
Split your ft: model ids into two groups by base.
- Base already on the deprecation list. The tune dies with its base. For the GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots that is 14 January 20271. The 6 January cutoff is almost irrelevant: a model you tune on the last day serves for about a week.
- Base with no announced shutdown. The tuned model keeps answering, but it is frozen. After 6 January you cannot refresh it with new labels, fix a regression in the training set or retrain after the base is updated. When OpenAI later announces that base's deprecation, you will be migrating without the option to retrain on the same platform.
Two other facts matter. OpenAI's notice policy promises at least six months for generally available models1, which gives you a floor for planning the second group. And you can read shutdown_date from GET /v1/models for each model1, so a nightly job can alarm on it instead of waiting for an email.
Do not count on a rescue. The OpenAI page describes no automatic upgrade for a retired id, and on Azure, fine-tuned workloads cannot be auto-upgraded and retirement dates are not extendable3.
Three paths, and what each one is for
| Path | Choose it when | Main risk |
|---|---|---|
| One last OpenAI job before 6 January | The base has no announced shutdown, and you need a refresh (new labels, a fixed training set) | The result is frozen, and the base's own deprecation arrives later with no retraining option on OpenAI |
| Tune elsewhere: Azure or an open-weight model | The behavior lives in the weights (labels, private knowledge) and prompting misses the pass mark | A new clock, new serving costs, new eval |
| Prompt a base model, with caching | The tune mostly fixed format, tone or a schema | Longer prompts; per-request cost rises with few-shot examples |
On "tune elsewhere", Microsoft's Foundry overview is the one vendor page I could read in full. It lists supervised fine-tuning for gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini and gpt-4.1-nano, reinforcement fine-tuning for o4-mini, and gpt-5 reinforcement fine-tuning by invitation only. It also lists open models with supervised fine-tuning: Llama-3.3-70B-Instruct, Qwen3-32B, gpt-oss-20b, Ministral-3B and several previews4. Microsoft's own caveat applies to the open ones: a method or tier that works for one model "isn't automatically supported for another", and an adapter "isn't interchangeable across similarly named base models"4.
Azure's clocks are later but still finite. Fine-tuned GPT-4.1-family models retire training no earlier than 14 April 2027 for existing customers, and deployment on 14 October 2027; for GPT-4o and GPT-4o mini the dates are 1 April and 1 October 20272. Tuning a GPT-4.1 on Azure today therefore buys months, not years. If the target is an open-weight base, the retirement risk moves from the vendor's deprecation page to your own upgrade plan.
Compare on cost per successful task
Per-token rates cannot choose between these paths, because a tune moves cost into three other places: training, hosting and the prompt. Use one metric across all candidates: monthly cost of everything that serves the task, divided by the tasks that pass the eval. The general method is in cost per successful task; the terms specific to this decision are these.
- Training, amortized over the months the tune will live. Azure bills supervised fine-tuning as training tokens × epochs × the training rate, and does not charge for queue time, failed jobs or jobs canceled before training starts5. Include reruns: Microsoft tells you to budget for repeated runs, grading and evaluation5.
- Hosting. Azure's Standard and Global Standard deployments of tuned models add an hourly hosting fee to token charges; Developer deployments have no hosting fee but are for evaluation only and are removed after 24 hours5. Managed compute bills GPU-hours, and allocated capacity "incurs charges while the deployment runs, even when you aren't sending inference requests"5.
- Inference, with the real prompt length. A tune lets you drop few-shot examples; prompting puts them back.
- Failures. Retries, human review of failed outputs and escalations count against the tasks that did not pass.
def cost_per_success(
inference: float, # monthly token spend on this path
hosting: float, # fixed monthly serving charge, 0 if none
one_off: float, # training + eval work for this path
months: int, # how long the path will live
tasks: int, # monthly task volume
pass_rate: float, # from the frozen eval, 0..1
) -> float:
monthly = inference + hosting + one_off / months
return monthly / (tasks * pass_rate)The open-weight row flips with volume. A GPU charge is fixed while token spend is linear, so a high, steady volume can favor dedicated hardware and a spiky one will not. Provider-side prompt caching rules differ in write price and lifetime; see prompt caching across providers before you assume a hit rate.
Build the eval before choosing
Everything above depends on the pass rate, and a pass rate is only as honest as the set it is measured on. Keep the tune's validation split, add sampled production requests including the ugly ones, and add known failures. Keep training rows out of it; Microsoft's guidance is to keep validation and final test examples out of the training set6, and its overview warns that benefits "aren't guaranteed" and that held-out evaluation is how you compare against a baseline4. Freeze the set and the pass mark before the first candidate runs.
A sensible order of work for the next seven weeks:
- This week: list every
ft:id, its base, its shutdown date and whether you still hold the training and validation files. - This week: if you lack recorded requests, turn on capture. It is not retroactive.
- Weeks 1 to 2: freeze the eval and run the untuned base with a prompt and schema. If it clears the mark, stop here.
- Weeks 2 to 4: if it misses by a small margin, add few-shot examples from the training set and a prompt cache; if by a lot, run a tune on an Azure or open-weight base.
- Before 6 January: decide on any last OpenAI job. Run one only if the base has no announced shutdown and a refresh would change the pass rate.
- Before the base's shutdown date: cut over with the old tune as the fallback until the last step. Fallback quality is its own test; see is your fallback model good enough.
Where a control plane helps
If your traffic already passes through INFRO on your own provider connections, the usage records group requests by model id, which finds every ft: caller, and the catalog records vendor shutdown dates and refuses a retired model before any provider call, naming the replacement. Datasets built from logged requests, experiments that run up to six variants over a dataset with a stated spend cap, and evaluators on your own account cover the comparison step; versioned prompts keep the few-shot variant rollback-able. INFRO sees request metadata, not answer quality, so the pass mark is still yours to set. It does not tune anything. The docs overview is at infro.io/docs (opens in a new tab)7.
The decision itself is small. Do not tune because you could until 6 January; tune only if a frozen eval shows prompting and caching cannot reach the mark, and then pick the platform whose retirement dates outlast the feature's useful life.
