Fine-Tuning

OpenAI is winding down fine-tuning: the platform is closed to new users, new job creation was restricted on Jul 2, 2026, and existing customers lose job creation on Jan 6, 2027. Fine-tuned models keep serving until their base models are retired, and no GPT-6 model can be fine-tuned.

OpenAI fine-tuning is being shut down: the platform "is no longer open to new users", new job creation was restricted on Jul 2, 2026, and existing customers can create training jobs only until Jan 6, 2027. Models you have already fine-tuned keep answering requests until their base model is deprecated. As of Nov 6, 2026 the only models the guide offers for tuning are GPT-4.1 and GPT-4.1 mini (plus GPT-4.1 nano, which has already been shut down) for supervised and preference tuning, and GPT-4o for vision tuning (fine-tuning guide); none of the GPT-6 models supports it (GPT-6.1 Sol). If you are starting a project now, plan on prompting, retrieval or open-weight models instead. For the rest of the platform, see OpenAI API.

Quick facts

Status	Winding down; closed to new users
Announced	May 7, 2026
New job creation restricted	Jul 2, 2026
Existing customers lose job creation	Jan 6, 2027
Inference on tuned models	Until the base model is deprecated
Tunable models	gpt-4.1, gpt-4.1-mini; gpt-4o for vision, as of Nov 6, 2026
Methods	Supervised, vision, direct preference optimization (DPO)
GPT-4.1 training price	$25.00 per 1M training tokens

The shutdown timeline

May 7, 2025: reinforcement fine-tuning launched (changelog).
Jun 4, 2025: direct preference optimization (DPO) launched (changelog).
May 7, 2026: OpenAI announced the end of self-serve fine-tuning (deprecations).
Jul 2, 2026: new job creation restricted.
Jan 6, 2027: existing customers can no longer create jobs.
Jan 14, 2027: GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down, with gpt-6-luna named as the replacement (deprecations).

The last two dates matter together. Inference on a fine-tuned model continues only "until their base models are deprecated" (pricing), and the base models for most tuned GPT-4.1 and GPT-4o models shut down on Jan 14, 2027. Read literally, a model tuned on gpt-4.1 or gpt-4o stops working that day. OpenAI's pages read for this article do not say anything more specific about tuned models, so check your own models' pages in the dashboard.

Which methods were offered

The fine-tuning guide still documents three methods as of Nov 6, 2026 (fine-tuning guide):

Method	What it does	Models
Supervised fine-tuning	Trains on example prompts and ideal answers	gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, gpt-4.1-nano-2025-04-14
Vision fine-tuning	Supervised training with images in the examples	gpt-4o-2024-08-06
Direct preference optimization (DPO)	Trains on pairs of a preferred and a rejected answer	The three gpt-4.1 models

OpenAI's pages disagree on one row: the guide still lists gpt-4.1-nano, but the deprecations page shut gpt-4.1-nano down on Oct 23, 2026, and the pricing page has no fine-tuning row for it. Treat nano as unavailable.

Reinforcement fine-tuning, launched on May 7, 2025, is no longer mentioned in the guide. Whether existing reinforcement fine-tuning jobs or models still run was not stated on any page read for this article.

Fine-tuning prices

Base model	Training per 1M tokens	Inference (in / cached / out) per 1M tokens
gpt-4.1-2025-04-14	$25.00	$3.00 / $0.75 / $12.00
gpt-4.1-mini-2025-04-14	$5.00	$0.80 / $0.20 / $3.20
gpt-4o-2024-08-06	$25.00	$3.75 / $1.875 / $15.00
gpt-4o-mini-2024-07-18	$3.00	$0.30 / $0.15 / $1.20

Source: OpenAI API pricing, read Nov 6, 2026.

Tuned models cost more to run than their base models: a tuned gpt-4.1 is $3.00 input and $12.00 output against $2.00 and $8.00 for plain gpt-4.1. OpenAI offers "inference discounts" if you enable data sharing when creating the job (pricing). Fine-tuning has its own rate limits, listed on the dashboard's limits page and through GET /v1/fine_tuning/model_limits (rate limits guide).

What to do instead

For new projects and for teams moving off a tuned model, the options with sources behind them are:

Move to GPT-6 Luna with good prompts. It is OpenAI's named replacement for GPT-4.1 and GPT-4o and costs $0.10 input and $0.50 output per 1M tokens, far less than a tuned gpt-4.1-mini. Put examples in the prompt and cache them, since cached input on Luna costs $0.01 per 1M tokens (pricing).
Use Structured Outputs when the reason for tuning was a fixed output format: it forces answers to match your JSON Schema (structured outputs guide). See Responses API.
Use file search when the reason was knowledge of your documents: it costs $2.50 per 1,000 calls plus storage.
Use Skills for repeatable procedures: versioned skill bundles in the hosted shell (tools guide).
Tune an open-weight model yourself. OpenAI's gpt-oss models carry an Apache 2.0 license (gpt-oss-120b), so you can train them on your own hardware.

Before moving, run your old test set against the new model. OpenAI's own Evals platform cannot help: it went read-only on Oct 31, 2026 and shuts down on Nov 30, 2026, and OpenAI points to Promptfoo instead (deprecations).

Frequently asked questions

Can I still fine-tune OpenAI models? Only if your organization already used fine-tuning, and only until Jan 6, 2027. OpenAI closed fine-tuning to new users and restricted new job creation on Jul 2, 2026.

Can I fine-tune GPT-6? No. OpenAI's model pages list fine-tuning as not supported for the GPT-6 models, and the fine-tuning guide covers only GPT-4.1 and GPT-4o, as of Nov 6, 2026.

What happens to my fine-tuned OpenAI model? It keeps serving requests until its base model is deprecated. GPT-4.1, GPT-4.1 mini, GPT-4o and GPT-4o mini shut down on Jan 14, 2027, so tuned models built on them should be replaced before then.

How much does OpenAI fine-tuning cost? As of Nov 6, 2026, training costs $25 per 1M training tokens on gpt-4.1 and gpt-4o, $5 on gpt-4.1-mini and $3 on gpt-4o-mini, and tuned models cost more per token to run than their base models.

What should I use instead of OpenAI fine-tuning? OpenAI names GPT-6 Luna as the replacement for the tunable GPT-4.1 and GPT-4o models; pair it with prompt examples, Structured Outputs or file search, or tune an open-weight model such as gpt-oss yourself.