Fine-tune shutdown checker
Will my OpenAI fine‑tune stop working?
Paste a fine-tuned model ID. You'll see the date it stops, what OpenAI names as its replacement, and the options we measured, with what each costs.
Dates checked 10 and 21 October 2026Runs in your browser
Scripts are off in this browser, so the checker can't run. Every base
model and its date is in the table below.
Result
Your three options
GPT-5.6 Luna, shown your most similar labelled examples
Tied a fine-tuned model
on two of three public tests: 93.8% against 94.0% on Banking77 and 94.6% against 95.1% on
CLINC150. Trailed by 2.0 points on long CFPB complaints, 87.5% against 89.5% (products merged
after seeing the errors). About $0.37 per 1,000 short messages and $1.20 per 1,000 long
complaints, from our OpenAI bill for 8–9 October, before the search for the examples. Set up
and tested on your data: $750 fixed.
An embeddings classifier you own
Tied on Banking77, 93.5%. Fell behind on
intents with off-topic requests (CLINC150, 90.6%) and on long complaints (CFPB, 86.2%). One CPU
server, about $58 a month by our estimate.
A small model you own
Qwen3-4B with LoRA: 94.0%, 95.1% and 89.5% on the same
three tests. One always-on L4 GPU with its server, about $815 a month by our estimate, flat up to
one GPU's capacity, about 130 million short messages a month. $3,000 fixed per task, refunded
in full if it misses the margin we agree.
Public datasets, not your data. The CFPB scores merge products the agency renamed,
a choice we made after seeing the errors. Every number, prompt and limit:
GPT-5.6 Luna vs a fine-tuned model.
This week
No data needed for the call. The check runs in your cloud account or on a
machine you control, so your examples never leave your infrastructure.
Base model:
Stops on 23 October 2026
That day OpenAI stops serving it, so expect calls to it to fail and the feature built on it to stop.
- OpenAI's named replacement
gpt-5.6-luna, which can't be fine-tuned
Two of the three replacements OpenAI has named for it since April 2026 now have shutdown dates of their own: gpt-5-nano on 11 December 2026 and gpt-5.4-nano on 1 April 2027.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
Nano was the cheap choice for short, high-volume answers such as labels and routes. OpenAI names GPT-5.6 Luna, the model we measured, so test it first.
Before 23 October 2026, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with. If it stops as scheduled, you can't compare a replacement against it afterwards (
how to save them).
Call it now. If it still answers, save its answers on real inputs today. If calls fail and you still have the file you trained it with, with labels from people or your records, the
48-hour rescue rebuilds the task as a model you own, paid only on acceptance.
Base model:
Stops on 23 October 2026
That day OpenAI stops serving it, so expect calls to it to fail and the feature built on it to stop.
- OpenAI's named replacement
gpt-5.6-terra, which can't be fine-tuned
o4-mini is the only model OpenAI lists for reinforcement fine-tuning, so on OpenAI's platform the job can't move to another model. Your graded examples and your grader can.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
If your grader checks a fixed answer, such as a label, a score or a field, the task fits the options below. If it judges open-ended reasoning, test a current reasoning model such as Terra first; we haven't measured Terra.
Before 23 October 2026, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the training and validation files and your grader's definition. If it stops as scheduled, you can't compare a replacement against it afterwards (
how to save them).
Call it now. If it still answers, save its answers on real inputs today. If calls fail and you still have the file you trained it with, with labels from people or your records, the
48-hour rescue rebuilds the task as a model you own, paid only on acceptance.
Base model:
Stops on 23 October 2026
That day OpenAI stops serving it, so expect calls to it to fail and the feature built on it to stop.
- OpenAI's named replacement
gpt-5.6-terra, which can't be fine-tuned
Some pages still name gpt-5.4-mini, from OpenAI's list as it stood in May 2026. For classification, extraction and routing, OpenAI's own GPT-5.6 upgrade guide points to gpt-5.6-luna instead.
OpenAI shut down the base gpt-3.5-turbo-1106 snapshot on 28 September 2026. It lists gpt-3.5-turbo fine-tunes for 23 October; if calls to yours already fail, see the 48-hour rescue.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
Most gpt-3.5-turbo fine-tunes classify, route, tag or pull fixed fields. For those, test GPT-5.6 Luna with your most similar labelled examples first.
Before 23 October 2026, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with. If it stops as scheduled, you can't compare a replacement against it afterwards (
how to save them).
Call it now. If it still answers, save its answers on real inputs today. If calls fail and you still have the file you trained it with, with labels from people or your records, the
48-hour rescue rebuilds the task as a model you own, paid only on acceptance.
Base model:
Stops on 23 October 2026
That day OpenAI stops serving it, so expect calls to it to fail and the feature built on it to stop.
- OpenAI's named replacement
gpt-5.6-sol, which can't be fine-tuned
Pages written from OpenAI's list as it stood in May 2026 still name gpt-5.5.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
If it does a narrow task with answers you can check, such as labels, routes or fields, these are the options we measured. Long-form writing usually doesn't move to a small model: test Sol or Terra with a prompt instead. We haven't measured either.
Before 23 October 2026, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with. If it stops as scheduled, you can't compare a replacement against it afterwards (
how to save them).
Call it now. If it still answers, save its answers on real inputs today. If calls fail and you still have the file you trained it with, with labels from people or your records, the
48-hour rescue rebuilds the task as a model you own, paid only on acceptance.
Base model:
Stops on 23 October 2026
That day OpenAI stops serving it, so expect calls to it to fail and the feature built on it to stop.
- OpenAI's named replacement
gpt-5.6-terra, which can't be fine-tuned
The base babbage-002 and davinci-002 models themselves were shut down on 28 September 2026, and new training on them ended on 28 October 2024.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
These were completion models, often trained as classifiers. Your prompt and completion file trains a model you own as it is, if its completions were written or checked by people or come from your records.
Before 23 October 2026, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with. If it stops as scheduled, you can't compare a replacement against it afterwards (
how to save them).
Call it now. If it still answers, save its answers on real inputs today. If calls fail and you still have the file you trained it with, with labels from people or your records, the
48-hour rescue rebuilds the task as a model you own, paid only on acceptance.
Base model:
Stops on 14 January 2027
OpenAI's notice of 14 October 2026 removes this base model on 14 January 2027, and fine-tunes built on it stop the same day. No one can start a fine-tuning job from 6 January 2027, so it can't be retrained before then either.
- OpenAI's named replacement
gpt-6-luna
Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
It stops on 14 January 2027 and can't be retrained after 6 January. These are the options we measured; test the cheapest first.
Nothing stops on 23 October for this one. Before 6 January 2027, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with, so you can test a replacement against it before 14 January (
how to save them).
Base model:
Stops on 14 January 2027
OpenAI's notice of 14 October 2026 removes this base model on 14 January 2027, and fine-tunes built on it stop the same day. No one can start a fine-tuning job from 6 January 2027, so it can't be retrained before then either.
- OpenAI's named replacement
gpt-6-luna
Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
It stops on 14 January 2027 and can't be retrained after 6 January. These are the options we measured; test the cheapest first.
Nothing stops on 23 October for this one. Before 6 January 2027, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with, so you can test a replacement against it before 14 January (
how to save them).
Base model:
Stops on 14 January 2027
OpenAI's notice of 14 October 2026 removes this base model on 14 January 2027, and fine-tunes built on it stop the same day. No one can start a fine-tuning job from 6 January 2027, so it can't be retrained before then either.
- OpenAI's named replacement
gpt-6-luna
Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
It stops on 14 January 2027 and can't be retrained after 6 January. These are the options we measured; test the cheapest first.
Nothing stops on 23 October for this one. Before 6 January 2027, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with, so you can test a replacement against it before 14 January (
how to save them).
Base model:
Stops on 14 January 2027
OpenAI's notice of 14 October 2026 removes this base model on 14 January 2027, and fine-tunes built on it stop the same day. No one can start a fine-tuning job from 6 January 2027, so it can't be retrained before then either.
- OpenAI's named replacement
gpt-6-luna
Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
It stops on 14 January 2027 and can't be retrained after 6 January. These are the options we measured; test the cheapest first.
Nothing stops on 23 October for this one. Before 6 January 2027, log this model's answers on 2,000–5,000 real inputs, with the model ID, prompt and settings, and keep the file you trained it with, so you can test a replacement against it before 14 January (
how to save them).
Base model:
Stopped on 4 January 2024
Models fine-tuned through OpenAI's first fine-tuning endpoint, /v1/fine-tunes, were shut down with it on 4 January 2024.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
It can't be measured any more, but the file you trained it with can still rebuild the task.
It can't be asked anything now, so there is nothing to log. If you still have the file you trained it with, and its labels came from people or your records, the
48-hour rescue rebuilds the task as a model you own, paid only on acceptance.
Base model:
On Microsoft's schedule, not OpenAI's
This looks like an Azure OpenAI fine-tune. OpenAI's 23 October date is for models on OpenAI's own platform.
Source: Azure model retirement schedule, read 10 October 2026.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
Azure buys time, then you move again. These are the options we measured.
Nothing stops on 23 October on Azure. Save a labelled test set with this model's answers on it now, so you can test a replacement before Microsoft's date.
Base model:
A job ID, not a model ID
IDs that start with ftjob- name the training job. The model it made is in the job's fine_tuned_model field and starts with ft:. Paste that one.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
Base model:
Not recognised
We don't recognise this base model. In a fine-tuned ID such as ft:gpt-4.1-nano-2025-04-14:your-org::8xYz1aBc, the base model is the part after ft:. Check yours against OpenAI’s deprecations page, or send it to us and we'll look.
There is no ft: prefix, so this looks like a base model, not a fine-tune. This is what happens to fine-tunes built on it.
Where to find your IDs: search your code and config for ft:, open the fine-tuning page of your OpenAI dashboard, or list your organization's models with GET /v1/models.
01 Dates
Every fine-tuned base model, and what happens to it
The base model is the part of the ID after ft:. In ft:gpt-3.5-turbo-0125:your-org::9fKe3Lw1 it is gpt-3.5-turbo.
| Base model in your ID | What happens to its fine-tunes | OpenAI's named replacement | More |
|---|
| gpt-4.1-nano-2025-04-14 | Stops on 23 October 2026Two of the three replacements OpenAI has named for it since April 2026 now have shutdown dates of their own: gpt-5-nano on 11 December 2026 and gpt-5.4-nano on 1 April 2027. | gpt-5.6-luna | Replacing a fine-tuned gpt-4.1-nano |
|---|
| o4-mini-2025-04-16, including reinforcement fine-tunes | Stops on 23 October 2026o4-mini is the only model OpenAI lists for reinforcement fine-tuning, so on OpenAI's platform the job can't move to another model. Your graded examples and your grader can. | gpt-5.6-terra | Replacing a fine-tuned o4-mini (RFT) |
|---|
| gpt-3.5-turbo (0125, 1106 or 0613) | Stops on 23 October 2026Some pages still name gpt-5.4-mini, from OpenAI's list as it stood in May 2026. For classification, extraction and routing, OpenAI's own GPT-5.6 upgrade guide points to gpt-5.6-luna instead. | gpt-5.6-terra | Replacing a fine-tuned gpt-3.5-turbo |
|---|
| gpt-4-0613 | Stops on 23 October 2026Pages written from OpenAI's list as it stood in May 2026 still name gpt-5.5. | gpt-5.6-sol | Replacing a fine-tuned gpt-4 |
|---|
| babbage-002 or davinci-002 | Stops on 23 October 2026The base babbage-002 and davinci-002 models themselves were shut down on 28 September 2026, and new training on them ended on 28 October 2024. | gpt-5.6-terra | The shutdown checklist: babbage-002 and davinci-002 |
|---|
| gpt-4.1-mini-2025-04-14 | Stops on 14 January 2027, with its base model; no new training jobs from 6 January 2027Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026. | gpt-6-luna | – |
|---|
| gpt-4.1-2025-04-14 | Stops on 14 January 2027, with its base model; no new training jobs from 6 January 2027Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026. | gpt-6-luna | – |
|---|
| gpt-4o-mini-2024-07-18 | Stops on 14 January 2027, with its base model; no new training jobs from 6 January 2027Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026. | gpt-6-luna | – |
|---|
| gpt-4o-2024-08-06 or gpt-4o-2024-11-20 | Stops on 14 January 2027, with its base model; no new training jobs from 6 January 2027Announced by OpenAI on 14 October 2026; read on its deprecations page on 21 October 2026. | gpt-6-luna | – |
|---|
ada, babbage, curie or davinci, in an old-style ID such as curie:ft-your-org-2023-03-01 | Stopped on 4 January 2024 | None named | – |
|---|
Read on 10 October 2026 from OpenAI's deprecations page (the notice of 22 April 2026 for the 23 October list, and the fine-tuning timeline of 7 May 2026), its model optimization guide for what can still be trained, and its pricing page, which still shows training prices for some models the guide no longer lists. The 14 January 2027 dates for gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini come from OpenAI's notice of 14 October 2026, read on the deprecations page on 21 October 2026. The 10 October reading was through two public copies of OpenAI's docs, which agreed; the links go to the live pages. GPT-5.6 Luna, Terra and Sol all list fine-tuning as "Not supported" on their model pages. Azure fine-tunes follow Microsoft's own schedule. Dates have moved before: check the source before you plan around one.
02 Options
Three options we measured, and what they cost
On three public classification tests, OpenAI's GPT-5.6 Luna, given the 20 most similar past examples, tied a small fine-tuned model on two and trailed by 2 points on the third. A plain embeddings classifier tied on one of them. The cheapest option that passes differs by task, so we test them all on your data before you pay.
| Option | Banking773,080 support messages | CLINC1505,500 requests, some off-topic | CFPB5,117 long complaints | Running cost |
|---|
| GPT-5.6 Luna, shown the 20 most similar labelled examples | 93.8%tied | 94.6%tied | 87.5%2.0 behind | About $0.37 per 1,000 short messages and $1.20 per 1,000 long complaints, from our OpenAI bill for 8–9 October |
| bge-base embeddings and a classifier, which you own | 93.5%tied | 90.6% | 86.2% | One CPU server: about $58 a month by our estimate |
| Qwen3-4B fine-tuned with LoRA, which you own | 94.0% | 95.1% | 89.5% | One always-on L4 GPU with its server: about $815 a month by our estimate, flat up to one GPU's capacity, about 130 million short messages a month |
- "Tied" means the 95% confidence interval of the paired difference from our fine-tuned model includes zero. We registered the Luna tests before running them; the embeddings classifier's scores come from our earlier bake-off.
- The CFPB scores merge products the agency renamed, a choice we made after seeing the errors.
- Off-topic requests on CLINC150 are a trade-off: our model caught more of them than Luna with examples, 83.9% against 79.1%, but wrongly turned away slightly more real requests, 0.40% against 0.18%.
- Luna's cost leaves out the search for the examples before each call, so read it as a floor. Our bill doesn't split by model.
- These are public datasets, not your data. Every number, prompt and limit: GPT-5.6 Luna vs a fine-tuned model.
Our fees. The test on your data is at no charge before you commit. If Luna passes, setting it up with retrieved examples and testing it on your data is $750 fixed. A model you own is $3,000 fixed per classification or routing task; the first three clients pay nothing up front and owe nothing unless the final model meets the margin we agree (pricing). Already stopped? The 48-hour rescue is $1,500, paid on acceptance.
If Luna passes on your data, use Luna. We'll tell you when it does. The owned model earns its place on long or messy text, at high volume, or when the cheaper options miss your bar.