Replacing a fine‑tuned gpt-3.5-turbo before 23 October 2026
On 23 October 2026 OpenAI stops serving fine-tunes of gpt-3.5-turbo, and the base gpt-3.5-turbo stops the same day. OpenAI's deprecations page names gpt-5.6-terra as the replacement base model, and Terra can't be fine-tuned. Pages written between May and July 2026 may still say gpt-5.4-mini: that was OpenAI's list then. Most gpt-3.5-turbo fine-tunes classify, route, tag or pull fixed fields, and for those jobs OpenAI's own GPT-5.6 upgrade guide points to gpt-5.6-luna, not Terra. Luna is the model we measured: given the 20 most similar labelled examples, it tied a small fine-tuned model on two of three public tests and trailed by 2 points on the third. The chat-format file you trained with also trains a model you own. Before 23 October, save your fine-tune's answers on real inputs, then test the options on your own labelled data. If Luna passes, use Luna.
Not sure which of your models stop? Paste the IDs into the shutdown checker. Already stopped answering? The 48-hour rescue rebuilds the task from your training file, if its labels were written or checked by people or come from your records. Every retiring model: the shutdown checklist.
The dated facts, with sources
Every row links to where it comes from. These pages have changed before, so open the source before you plan around a date.
| Date | What happens | Source |
|---|---|---|
| 7 May and 2 July 2026 | Organizations that never fine-tuned can't create fine-tuning jobs (7 May). Neither can organizations with no inference on a fine-tuned model in the past 60 days (2 July). | OpenAI deprecations |
| May 2026 | A copy of OpenAI's deprecations page from May names gpt-5.4-mini as the replacement for fine-tuned gpt-3.5-turbo. Some third-party pages still repeat it. | That copy, in a public archive of OpenAI's docs |
| 28 September 2026 | The base snapshots gpt-3.5-turbo-1106 and gpt-3.5-turbo-instruct shut down (notice of 26 September 2025). This is a base-model date, not the fine-tune row. | OpenAI deprecations |
| As of 10 October 2026 | The "Recommended replacement base model" for ft-gpt-3.5-turbo is gpt-5.6-terra. The row names no snapshot. | OpenAI deprecations (notice of 22 April 2026) |
| As of 10 October 2026 | The gpt-5.6-terra model page lists fine-tuning as "Not supported". Terra "roughly corresponds to the mini model tier" and costs $2 per million input tokens and $12 per million output tokens. | GPT-5.6 Terra model page |
| As of 10 October 2026 | OpenAI's GPT-5.6 upgrade guide maps a "Nano model, classification, extraction, routing, high-volume, or strict-latency route" to gpt-5.6-luna, and a "Mini model, balanced lower-cost route" to gpt-5.6-terra. | Upgrading to GPT-5.6 Sol |
| As of 10 October 2026 | Supervised fine-tuning is listed only for gpt-4.1, gpt-4.1-mini and gpt-4.1-nano (2025-04-14 snapshots). gpt-3.5-turbo and gpt-4o-mini are no longer listed. OpenAI's pricing page still shows training prices for gpt-4o-mini-2024-07-18 and gpt-3.5-turbo (legacy), so the two pages disagree. | OpenAI model optimization guide and pricing page |
| 23 October 2026 | Fine-tunes of gpt-3.5-turbo stop. The base gpt-3.5-turbo-0125, and the gpt-3.5-turbo alias, stop the same day. | OpenAI deprecations (notice of 22 April 2026) |
| 6 January 2027 | "Active existing customers will no longer be able to create new fine-tuning jobs." | OpenAI deprecations |
| 14 January 2027 | OpenAI removes gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini from the API, with gpt-6-luna as the replacement. "Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models." | OpenAI deprecations (notice of 14 October 2026) |
Read on 10 October 2026 through two public copies of OpenAI's docs, which agreed. The 14 January 2027 row was read on OpenAI's deprecations page on 21 October 2026. The links go to the live pages, except the May copy.
Terra is the name on the list. Your task decides what to test.
Terra is OpenAI's mid-priced model. For labels and routes, OpenAI's own upgrade guide points to Luna, which costs a tenth as much per token ($0.20 against $2 per million input tokens, $1.20 against $12 output, list prices). Test Luna first on a label task, and Terra only if Luna misses your margin.
| What your fine-tune does | Test first | What we measured |
|---|---|---|
| Picks a label from a fixed list: classify, route, tag, triage | GPT-5.6 Luna shown your most similar labelled examples, an embeddings classifier, and a small model you own | Luna tied a fine-tuned 4B model on two of three public tests; the embeddings classifier tied on one (section 03) |
| Pulls fixed fields into JSON | Luna or Terra with a strict JSON schema, and a small model you own | Not measured by us. Our fixed fee covers classification and routing; extraction is quoted after the test. |
| Writes replies in a set tone or format | Terra or Luna with your best examples in the prompt, judged by people reading the replies side by side | Not measured. A small model can learn a format, but judge it on your own replies. |
What we measured on label tasks
Three public test sets; the Luna tests were pre-registered. We have never scored a fine-tuned gpt-3.5-turbo, so the comparison is with a small fine-tuned model of ours. Your own fine-tune's saved answers are the comparison that matters.
| Option | Banking77 3,080 messages | CLINC150 5,500 requests | CFPB 5,117 complaints |
|---|---|---|---|
| Qwen3-4B fine-tuned with LoRA, which you own | 94.0% | 95.1% | 89.5% |
| GPT-5.6 Luna, the 20 most similar labelled examples in each prompt | 93.8% | 94.6% | 87.5% |
| bge-base embeddings and a classifier, which you own | 93.5% | 90.6% | 86.2% |
| GPT-5.6 Luna, the category list only | 85.4% | 88.4% | 86.6% |
- The examples did the work. On Banking77 the same Luna scored 85.4% with the category list only, 87.2% with one fixed example of every category and 93.8% with the 20 most similar examples. A gpt-3.5-turbo fine-tune learned from your training file; Luna is shown the most similar rows of the same file for each message.
- Ties and gaps. Luna with examples tied our model on Banking77 (paired difference +0.2 points, 95% CI −0.6 to +0.9) and CLINC150 (+0.5, CI −0.04 to +1.0), and trailed by 2.0 points on long CFPB complaints (CI +1.2 to +2.9). The CFPB scores merge products the agency renamed, a choice we made after seeing the errors.
- Off-topic requests are a trade-off. On CLINC150 our model caught more of them, 83.9% against 79.1%, but wrongly turned away slightly more real requests, 0.40% against 0.18%.
Every table, the prompts and the limits: GPT-5.6 Luna vs a fine-tuned model.
The file you trained gpt-3.5-turbo with still works
The examples it retrieves
Each labelled row in your chat-format JSONL becomes an example. Before each call, the most similar rows are found with an embeddings model and sent with the request.
The training data, as it is
The same messages format trains a small open model whose weights you keep, or an embeddings classifier for simple triage. Hold out a test set first.
Where the answers came from
If people wrote or checked them, or they come from your records, they are labels. If a GPT model wrote them, tell us: we train only on labels people or your records confirmed, and we'll quote a labelling job or decline.
Find your own copy of the JSONL you uploaded, in your repository, bucket or notebook output, along with the system prompt and label list you used. Don't count on getting it back from OpenAI later.
What each option costs to run
| Option | Running cost | Our fee |
|---|---|---|
| GPT-5.6 Terra with a prompt | $2 per million input tokens, $12 per million output (list price; we haven't run it) | – |
| GPT-5.6 Luna with retrieved examples | 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 | $750 fixed to set up and test on your data |
| An embeddings classifier you own | One CPU server, about $58 a month by our estimate | Quoted after the test |
| A small model you own | 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; or your own cloud | $3,000 fixed per task, refunded in full if it misses the agreed margin |
The test on your data is at no charge before you commit (pricing). At low volume Luna is usually cheaper than a GPU; the test prices your real volume.
What to do this week
- Search your code, config and jobs for
ft:gpt-3.5-turbo, every snapshot: 0125, 1106 and 0613 fine-tunes all fall under the same row. - If one is built on gpt-3.5-turbo-1106, call it today. Its base snapshot shut down on 28 September 2026; if it already fails, go straight to the 48-hour rescue.
- Log each fine-tune's answers on 2,000–5,000 real inputs, with the model ID, prompt, settings and raw output, and keep logging live traffic until 23 October (how to save them).
- Label a test set of at least 1,000 real inputs, with labels from people or your records, not from the fine-tune.
- Run Luna with your most similar labelled examples on it first, then the other options, against a margin agreed before you look. If Luna passes, use Luna.
Common questions
What replaces a fine-tuned gpt-3.5-turbo: gpt-5.4-mini or gpt-5.6-terra?
On 10 October 2026, OpenAI's deprecations page names gpt-5.6-terra as the recommended replacement base model for ft-gpt-3.5-turbo. gpt-5.4-mini was the name in May 2026, and some pages still repeat it. Neither can be fine-tuned, so either way the replacement is a prompt, not a new fine-tune. For classification and routing, OpenAI's GPT-5.6 upgrade guide points to gpt-5.6-luna (OpenAI deprecations, read through public copies).
Can I fine-tune GPT-5.6 Terra or another GPT-5 model instead?
No. The GPT-5.6 Terra, Luna and Sol model pages all list fine-tuning as "Not supported" (Terra model page, checked 10 October 2026). With a GPT-5 model, your training data goes into the prompt instead: the most similar labelled examples, retrieved for each request.
Can I move my gpt-3.5-turbo fine-tune to gpt-4o-mini or gpt-4.1-mini on OpenAI?
Not as a destination. OpenAI's notice of 14 October 2026 removes gpt-4o-mini and gpt-4.1-mini, and the fine-tunes built on them, on 14 January 2027 (OpenAI deprecations, read 21 October 2026), and no one can start a fine-tuning job from 6 January 2027. gpt-4.1-nano is worse: its fine-tunes stop on 23 October 2026. Test GPT-5.6 Luna with your most similar labelled examples, and a model you own, from your own copy of the training file.
Does a gpt-3.5-turbo-1106 fine-tune stop on 23 October, or did it stop in September?
OpenAI's table lists gpt-3.5-turbo fine-tunes for 23 October 2026 without naming a snapshot, and the base gpt-3.5-turbo-1106 snapshot shut down on 28 September 2026. OpenAI also says inference on fine-tuned models is disabled when the underlying base model is deprecated, so the two lines don't settle it. Call yours today: if it answers, plan for 23 October; if it fails, the 48-hour rescue rebuilds it from your training file.
Is GPT-5.6 Luna good enough to replace a fine-tuned gpt-3.5-turbo classifier?
Often, and it's the cheapest option to test first. Given the 20 most similar labelled examples, it tied our fine-tuned 4B model on Banking77 (93.8% against 94.0%) and CLINC150 (94.6% against 95.1%) and trailed by 2.0 points on long CFPB complaints, with renamed products merged after seeing the errors. We have never scored a fine-tuned gpt-3.5-turbo, so test Luna against your own fine-tune's saved answers and your labels. If it meets your margin, use it.
Can I reuse my gpt-3.5-turbo training file?
Yes. The chat-format JSONL is the examples Luna retrieves from, and the training data for a model you own. Use it only where the answers were written or checked by people or come from your records; a model trained on another model's answers copies its mistakes.
Want the options tested on your data?
Start with a 20-minute call or a short email; no data needed. Then the test runs where your labelled examples and your fine-tune's saved answers already are: inside your cloud account or on a machine you control (how to prepare the file). Nothing leaves your infrastructure. Within 2 business days of getting access you get a straight answer, including "use Luna" when Luna passes.