Replacing a reinforcement fine‑tuned o4-mini before 23 October 2026
On 23 October 2026 OpenAI stops serving fine-tunes of o4-mini-2025-04-16, and the base o4-mini stops the same day. OpenAI's fine-tuning guide lists o4-mini for reinforcement fine-tuning (RFT) and no other model, and the substitute OpenAI names, gpt-5.6-terra, can't be fine-tuned. So on OpenAI's platform the training job can't move to another model. What moves is your graded data and your grader. If the grader checks a fixed answer, such as a label, a category or a field, the task is classification or extraction: test GPT-5.6 Luna with your most similar labelled examples, which tied a small fine-tuned model on two of three public classification tests, next to a small model you own trained on your reference answers. If the grader judges open-ended reasoning, test Terra or another current reasoning model with a prompt, scored by the same grader. We haven't measured Terra. Save your grader and your fine-tune's answers before 23 October.
Check every model ID you run in the shutdown checker. Other retiring models: the shutdown checklist. Already stopped answering? The 48-hour rescue fits if your grader checked a short, fixed answer.
The dated facts, with sources
Two shutdowns matter here: the model on 23 October, and OpenAI's Evals platform in October and November, whose graders OpenAI says are part of that transition.
| Date | What happens | Source |
|---|---|---|
| May 2026 | A copy of OpenAI's deprecations page from May names gpt-5.4-mini as the substitute for fine-tuned o4-mini. Some third-party pages 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 | ft-o4-mini-2025-04-16 is listed in the notice's main table, whose column is "Substitute model", with gpt-5.6-terra. It isn't in the table of fine-tuned models below it. | OpenAI deprecations (notice of 22 April 2026) |
| As of 10 October 2026 | Reinforcement fine-tuning is listed for o4-mini-2025-04-16 only, "Reasoning models only". Supervised fine-tuning is listed for gpt-4.1, gpt-4.1-mini and gpt-4.1-nano. | OpenAI model optimization guide |
| As of 10 October 2026 | The gpt-5.6-terra model page lists fine-tuning as "Not supported". Reasoning effort can be none, low, medium (the default), high, xhigh or max. $2 per million input tokens, $12 per million output. | GPT-5.6 Terra model page |
| 23 October 2026 | Fine-tuned o4-mini stops, and so do the base o4-mini-2025-04-16 and the o4-mini alias. A new o4-mini RFT job would stop the same day. | OpenAI deprecations (notice of 22 April 2026) |
| 31 October 2026 | Existing evals on OpenAI's Evals platform become read-only. "Graders documented for eval workflows are part of this transition." | OpenAI deprecations (notice of 3 June 2026) |
| 19 November 2026 | Azure retires its base o4-mini, listed as Deprecated (closed to new customers), with gpt-5.6-terra as the replacement. | Azure model retirement schedule |
| 30 November 2026 | The Evals dashboard and API are scheduled to shut down. | OpenAI deprecations (notice of 3 June 2026) |
| 6 January 2027 | No new fine-tuning jobs for anyone on OpenAI. | OpenAI deprecations |
| 2027 | On Azure, for existing customers only: fine-tuned o4-mini training stays open until no earlier than 19 May 2027, and deployments retire on 19 November 2027. Microsoft's lifecycle page gives 16 April and 16 October 2027 instead, so plan for the earlier dates. | Azure model retirement schedule and lifecycle page |
Read on 10 October 2026 through two public copies of OpenAI's docs, which agreed, and Microsoft's own docs repository. The links go to the live pages, except the May copy.
The weights stay. Your data and your grader move.
The fine-tuned weights
OpenAI doesn't release the weights of models trained on its platform, and no other model on OpenAI's own platform accepts an RFT job.
Your graded dataset
The prompts, and the reference answers your grader compared against. Where people or your records wrote those answers, they are labels: training data for a model you own and examples for a prompted one.
Your grader
It defined what a good answer is. It becomes the test every replacement has to pass, run on the same held-out prompts as your fine-tune's saved answers.
Pick the replacement by what your grader checks
OpenAI's guide pitches RFT at complex tasks that need reasoning, but many graders check a fixed answer. Those tasks are classification or extraction with a reasoning model in front, and cheaper options may pass.
| Your grader | What the task is | Test first | What we measured |
|---|---|---|---|
A string_check or other exact match against a reference label | Classification | GPT-5.6 Luna with your most similar labelled examples, an embeddings classifier, and a small model you own trained on the reference labels | Luna tied a fine-tuned 4B model on two of three public tests (section 04) |
| Checks fields of a JSON answer against reference values | Extraction to a schema | Luna or Terra with a strict JSON schema, and a small model you own | Not measured by us |
A score_model grader judging open-ended text or reasoning | Open-ended reasoning | Terra, or Sol for harder cases, with a prompt, at the lowest reasoning effort that passes | Not measured. A small model you own is usually the wrong tool here, and we'll say so. |
Rewrite the grader as code you run. o4-mini is scheduled to stop on 23 October, and OpenAI says graders documented for eval workflows are part of the Evals platform's shutdown: read-only from 31 October 2026, closed on 30 November 2026. A string check is a few lines of code; a score_model grader is a prompt you can send to any model. For label-type graders, the open-source parity check compares a candidate with your fine-tune's saved answers on the same items and returns PASS, FAIL or INCONCLUSIVE against a margin you set in advance.
What we measured, and where it applies
Only to graders that check a label. We have never scored an o4-mini RFT model or GPT-5.6 Terra. These are three public classification test sets; the Luna tests were pre-registered.
| 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% |
Luna with examples tied our model on Banking77 and CLINC150 and trailed by 2.0 points on CFPB, where the scores merge products the agency renamed, a choice we made after seeing the errors. GPT-5.6 Luna ran with reasoning effort "none"; "low" didn't raise accuracy on Banking77. On CLINC150's off-topic requests ours caught more, 83.9% against 79.1%, but wrongly turned away slightly more real requests, 0.40% against 0.18%. Every number: GPT-5.6 Luna vs a fine-tuned model.
Options to replace a fine-tuned o4-mini, and what they cost
| Option | Works after 23 October 2026 | Retrain it | Running cost |
|---|---|---|---|
| GPT-5.6 Terra with a prompt (OpenAI's substitute) | Yes | Can't be fine-tuned | $2 per million input tokens, $12 per million output, list price. Higher reasoning effort means more output tokens. Not run by us. |
| GPT-5.6 Luna with your most similar labelled examples (label tasks) | Yes | Can't be fine-tuned | 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 |
| Fine-tuned o4-mini on Azure (only if your subscription already deployed o4-mini) | Until late 2027 | Existing customers, until no earlier than April or May 2027 | Not checked for o4-mini |
| An embeddings classifier you own (labels only) | Yes | Yes, any time | One CPU server, about $58 a month by our estimate |
| A small model you own, trained on your reference answers (labels and fields) | Yes | Yes, any time | 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 |
Our fees: the test on your data is at no charge before you commit; Luna set up with retrieved examples and tested on your data is $750 fixed; a model you own is $3,000 fixed per classification or routing task, refunded in full if it misses the margin we agree; extraction is quoted after the test (pricing).
What to save this week
If it stops on 23 October as scheduled, you can't ask your fine-tune anything afterwards, and from 31 October any evals you keep on OpenAI are read-only.
- For each RFT job, copy the grader, the response schema and the settings. The fine-tuning jobs API returns them with the job.
- Keep your own copies of the training and validation JSONL.
- Run your fine-tune on 2,000–5,000 held-out prompts and save each answer with its grader score, the model ID and the settings.
- Rewrite the grader as code you run, and check that it gives the same scores as OpenAI's on a few hundred saved answers.
- Then score each candidate with it on the same prompts, against a margin agreed before you look.
Common questions
Does the 23 October 2026 shutdown include reinforcement fine-tuned o4-mini models?
Yes. OpenAI lists ft-o4-mini-2025-04-16 for shutdown on 23 October 2026, with gpt-5.6-terra as the substitute, and its fine-tuning guide lists o4-mini-2025-04-16 as the only model for reinforcement fine-tuning (OpenAI deprecations, read 10 October 2026 through public copies).
Can I run reinforcement fine-tuning on GPT-5.6 Terra or another newer OpenAI model?
Not on OpenAI. Terra's model page lists fine-tuning as "Not supported", as do the Luna and Sol pages, and OpenAI's guide lists no other model for reinforcement fine-tuning. From 6 January 2027 no fine-tuning jobs run on OpenAI at all. With Terra, what RFT taught the model has to come from the prompt instead.
What replaces a fine-tuned o4-mini: gpt-5.4-mini or gpt-5.6-terra?
On 10 October 2026 OpenAI's deprecations page names gpt-5.6-terra. gpt-5.4-mini was the name in May 2026, and some pages still repeat it. Neither can be fine-tuned. If your grader checks a label, also test GPT-5.6 Luna with your most similar labelled examples: for classification, OpenAI's own GPT-5.6 upgrade guide points to Luna.
Will my RFT grader still work after OpenAI's Evals platform shuts down?
Don't count on it. Existing evals become read-only on 31 October 2026 and the Evals dashboard and API shut down on 30 November 2026, and OpenAI says graders documented for eval workflows are part of that transition; it covers fine-tuning timelines separately. Either way, save the grader's definition from your RFT jobs now and run it as your own code, so it keeps working as the test for any replacement.
Can I keep o4-mini RFT running on Azure?
Only if your Azure subscription has already deployed o4-mini: Microsoft lists it as Deprecated, closed to new customers. For existing customers, its retirement schedule keeps fine-tuned o4-mini deployments until 19 November 2027, while its lifecycle page says 16 October 2027. You would retrain there, because OpenAI's weights don't transfer, and move again in 2027. Microsoft also lists gpt-5 (2025-08-07) for reinforcement fine-tuning, by invitation through your Microsoft account team; that base model retires on Azure on 9 February 2027 (retirement schedule, read 10 October 2026).
Can a small model I own replace an o4-mini reinforcement fine-tune?
When the grader checks a fixed answer, often: on three public classification tests our fine-tuned 4B model scored 94.0%, 95.1% and 89.5% (the last with renamed products merged, a choice we made after seeing the errors), and GPT-5.6 Luna with retrieved examples tied it on two. When the grader judges open-ended reasoning, usually not, and we'll say so. Test it either way with your own grader on held-out prompts.
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 graded prompts, their reference answers 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, and "not a fit" when the task is open-ended reasoning.