Replacing a fine‑tuned gpt-4 before 23 October 2026
On 23 October 2026 OpenAI stops serving fine-tunes of gpt-4 (gpt-4-0613), and the base gpt-4 stops the same day. OpenAI's deprecations page names gpt-5.6-sol as the replacement base model, and Sol can't be fine-tuned. Pages written from OpenAI's May 2026 list still say gpt-5.5. What to test depends on the job. For long-form writing or hard reasoning, test Sol, or the cheaper Terra, with your best training examples in the prompt; we haven't measured either, and a small model is rarely the answer there. For a narrow task with answers you can check, such as labels, routes or fields, a much smaller model may be enough: GPT-5.6 Luna, given the 20 most similar labelled examples, tied a small fine-tuned model on two of three public classification tests. Retraining on gpt-4.1 doesn't buy much time: OpenAI's notice of 14 October 2026 removes gpt-4.1 and the fine-tunes built on it on 14 January 2027.
Check every model ID you run in the shutdown checker. Other retiring models: the shutdown checklist.
The dated facts, with sources
| Date | What happens | Source |
|---|---|---|
| May 2026 | A copy of OpenAI's deprecations page from May names gpt-5.5 as the replacement for fine-tuned gpt-4. Pages and AI answers written then still repeat it. | That copy, in a public archive of OpenAI's docs |
| 7 May and 2 July 2026 | Organizations that never fine-tuned can't create fine-tuning jobs (7 May), and neither can organizations with no inference on a fine-tuned model in the past 60 days (2 July). | OpenAI deprecations |
| As of 10 October 2026 | The "Recommended replacement base model" for ft-gpt-4 is gpt-5.6-sol. | OpenAI deprecations (notice of 22 April 2026) |
| As of 10 October 2026 | The gpt-5.6-sol model page lists fine-tuning as "Not supported". Sol costs $4 per million input tokens and $20 per million output, a promotional price "available at least through November 21, 2026". | GPT-5.6 Sol model page |
| As of 10 October 2026 | OpenAI's GPT-5.6 upgrade guide warns against "choosing Sol for workloads that were intentionally mini, nano, low-cost, or latency-sensitive", and maps classification, extraction and routing to gpt-5.6-luna. | Upgrading to GPT-5.6 Sol |
| As of 10 October 2026 | Supervised fine-tuning is listed for gpt-4.1, gpt-4.1-mini and gpt-4.1-nano (2025-04-14 snapshots). | OpenAI model optimization guide and deprecations |
| 23 October 2026 | Fine-tunes of gpt-4 stop. The base gpt-4-0613, and the gpt-4 alias, stop the same day. | OpenAI deprecations (notice of 22 April 2026) |
| 6 January 2027 | No new fine-tuning jobs for anyone, so no fine-tune can be retrained after that. | OpenAI deprecations |
| 14 January 2027 | OpenAI removes gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini from the API. "Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models", so a gpt-4.1 fine-tune made now stops that day. The named replacement is gpt-6-luna. | 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.
What was your gpt-4 fine-tune for?
Which replacement fits depends on what the fine-tune did. Some tasks still need a large model; many narrow ones don't.
| The task | Test first | What we know |
|---|---|---|
| Long-form writing in a house style: reports, replies, summaries | Sol or Terra with your best training examples in the prompt, judged by people reading the outputs side by side with your fine-tune's saved ones | Not measured. A small model you own rarely matches this, and we'll say so. |
| Hard reasoning or analysis with a checkable result | Sol at a few reasoning efforts, scored against your labelled answers | Not measured |
| Labels, routes or fixed fields | GPT-5.6 Luna with 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 classification tests (section 03) |
Your training file goes into the prompt
Sol can't learn from your file the way gpt-4 did, but it can be shown the right part of it. Index your labelled examples, and before each request retrieve the ones most similar to it.
We measured this on a smaller model, GPT-5.6 Luna, in tests we pre-registered on three public classification sets. On Banking77 (3,080 support messages), Luna scored 87.2% with one fixed example of every category and 93.8% with the 20 most similar examples, against 94.0% for our fine-tuned 4B model: the examples did the work. With the same setup it tied on CLINC150 (94.6% against 95.1%) and trailed by 2.0 points on long CFPB complaints (87.5% against 89.5%, with renamed products merged after seeing the errors). We didn't test Sol, Terra or any writing task. Every number: GPT-5.6 Luna vs a fine-tuned model.
Twenty examples in every prompt add up
On short Banking77 messages, a prompt with 20 retrieved examples was about 1,600 input tokens. At current prices ($4 against $0.20 per million input tokens), that input costs 20 times as much on Sol as on Luna.
| Option | Price | Notes |
|---|---|---|
| GPT-5.6 Sol (OpenAI's named replacement) | $4 per million input tokens, $20 output | Promotional price, at least through 21 November 2026. Not run by us. |
| GPT-5.6 Terra | $2 per million input tokens, $12 output | List price. Not run by us. |
| GPT-5.6 Luna with retrieved examples | About $0.37 per 1,000 short messages, $1.20 per 1,000 long complaints | From our OpenAI bill for 8–9 October, before the search for the examples. List price $0.20 input, $1.20 output. |
| A small model you own (narrow tasks) | About $815 a month for one always-on L4 GPU with its server, by our estimate | Flat up to one GPU's capacity, about 130 million short messages a month. Our fee: $3,000 fixed per classification or routing task, refunded in full if it misses the agreed margin. |
Model prices from OpenAI's model pages, read 10 October 2026. The test on your data is at no charge before you commit; if Luna passes, setting it up is $750 fixed (pricing).
What to do this week
- Search your code and config for
ft:gpt-4-0613andft:gpt-4:, and paste every ID into the checker. - Save the fine-tune's answers on 2,000–5,000 real inputs, with the model ID, prompt, settings and raw output (how to save them). For writing tasks, these saved outputs are what reviewers compare a replacement against.
- Keep your own copy of the training JSONL and the system prompt.
- Test the replacement on held-out inputs before you switch: labels against a margin agreed in advance, writing by people reading both outputs blind.
Common questions
What replaces a fine-tuned gpt-4: gpt-5.5 or gpt-5.6-sol?
On 10 October 2026 OpenAI's deprecations page names gpt-5.6-sol as the recommended replacement base model for ft-gpt-4. gpt-5.5 was the name in May 2026, and pages written then still repeat it. Neither can be fine-tuned, so the replacement is a prompt, not a new fine-tune (OpenAI deprecations, read through public copies).
Can I fine-tune GPT-5.6 Sol?
No. Sol's model page lists fine-tuning as "Not supported", and so do the Terra and Luna pages (Sol model page, checked 10 October 2026). Put the most similar examples from your training file into each prompt instead.
Can I retrain my gpt-4 fine-tune on gpt-4.1?
You may be able to, but it doesn't buy much time. OpenAI's notice of 14 October 2026 removes gpt-4.1 and the fine-tunes built on it on 14 January 2027, with gpt-6-luna as the named replacement, and no one can start a fine-tuning job from 6 January 2027 (OpenAI deprecations, read 21 October 2026). Organizations that never fine-tuned have been refused since 7 May 2026; ours was, on 7 October.
Will a smaller model do what my fine-tuned gpt-4 did?
For narrow tasks with checkable answers, often: on three public classification tests, GPT-5.6 Luna with the 20 most similar labelled examples tied a small fine-tuned model on two and trailed by 2.0 points on the third, long complaints with renamed products merged after seeing the errors. For long-form writing, usually not, and we'll say so. Test on your own inputs against your fine-tune's saved answers before you decide.
How long will Sol's current price last?
OpenAI's Sol model page says its promotional pricing of $4 per million input tokens and $20 per million output is "available at least through November 21, 2026" (checked 10 October 2026). It doesn't say what comes after, so budget for a change and recheck the page.
Not sure which kind of task yours is?
Email us, or ask for a 20-minute call; no data needed. Then the test runs on your inputs and the answers you'd accept, where they 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 "not a fit for a small model" when it isn't.