Unlocking Precision with OpenAI Fine-tuning: OpenAI Fine tuning

If you are evaluating OpenAI fine-tuning for AI agents, here is the direct answer: fine-tuning trains a base GPT model on your own examples so it reliably speaks your industry's language, follows your formats, and holds your tone — and in 2026 it comes with an expiry date, because OpenAI is winding the self-serve platform down. This guide explains what fine-tuning is, the four methods available, what it costs, how to run a job step by step, and when it is the wrong tool entirely.

What OpenAI fine-tuning actually is

OpenAI's pre-trained models are generalists: they learned from enormous amounts of public text and handle a wide range of tasks reasonably well. Fine-tuning takes one of these base models and trains it further on a smaller, curated dataset of your examples. The process adjusts the model's weights so it learns to:

• Recognize domain context. Industry terminology, internal jargon, product names, and the nuances of your field.
• Hold a style or tone. Your brand voice — formal, empathetic, or terse — applied consistently across thousands of responses.
• Follow a fixed format. Classification taxonomies, extraction schemas, and report structures reproduced exactly, every time.
• Fail less on specific tasks. Fewer instruction-following errors on the narrow task you trained for.

Fine-tuning is fundamentally different from prompt engineering. A prompt steers behavior for a single call; fine-tuning changes the model itself, so the behavior persists without repeating instructions in every request. What fine-tuning does not do is add fresh factual knowledge — that is the job of retrieval-augmented generation (RAG), covered below.

The fine-tuning methods available

OpenAI's platform supports four fine-tuning methods, and they solve different problems (OpenAI cookbook):

• Supervised fine-tuning (SFT). You provide input/output pairs that demonstrate the behavior you want, and the model learns to imitate them. SFT is the default and the right choice for classification, structured extraction, format enforcement, and tone. It is available on the GPT-4.1 family (gpt-4.1, gpt-4.1-mini, gpt-4.1-nano) and earlier GPT-4o snapshots (OpenAI fine-tuning docs).
• Vision fine-tuning. SFT extended to image-and-text pairs, for domain-specific visual tasks such as document image classification or visual question answering.
• Direct preference optimization (DPO). Instead of perfect answers, you provide pairs — a preferred response and a rejected one — and the model shifts toward the preferred pattern. DPO suits subjective quality: tone, politeness, brand voice. A beta hyperparameter between 0 and 2 controls how aggressively the model adapts, and OpenAI recommends running SFT first, then DPO on top of that checkpoint.
• Reinforcement fine-tuning (RFT). You supply prompts plus a programmatic grader — a unit test, a regex, a scoring function — and the model trains to maximize the grader's score. RFT is available on the o4-mini reasoning model and suits tasks with verifiable correctness: math, code, structured reasoning.

The GPT-5 family is not available for fine-tuning (mlq.ai's guide to what replaced GPT-3 fine-tuning).

The 2026 wind-down: read this before you start

In May 2026 OpenAI announced it is winding down the self-serve fine-tuning platform. New organizations can no longer create fine-tuning jobs, and existing customers can start new training runs until 6 January 2027, according to OpenAI's published timeline (mlq.ai). Fine-tuned models already in production keep serving inference until their underlying base model is deprecated (OpenAI pricing docs).

Practical consequences:

• If you already run fine-tuned models, they keep working, but plan for retraining to stop. Freeze your best checkpoints, keep training data and eval sets versioned, and map a fallback path to a base model plus retrieval.
• If you were about to start, the OpenAI platform is closed to you. The realistic alternatives are stronger prompting on a frontier base model, RAG for knowledge, structured outputs for format, and — where custom weights are truly required — fine-tuning an open-weight model on a hosted platform instead.
• Either way, OpenAI is steering users toward better prompting, retrieval, tool use, and eval-driven iteration rather than custom weights.

When fine-tuning beats prompting and RAG — and when it is the wrong choice

The decision order practitioners recommend (Respan's fine-tuning guide):

• Is the task fact retrieval? Use RAG, not fine-tuning. A tuned model bakes in patterns, not knowledge, and will confidently serve stale facts. Our RAG or fine-tuning decision guide for banks and insurers walks through this trade-off for regulated industries.
• Is the failure tone or format? Try a stricter prompt and a structured output schema first. Fine-tuning is overkill for "always answer in JSON."
• Can you write a rubric for what good looks like? If not, you cannot evaluate the fine-tune, and you should not start one.
• Do you have at least 50–100 high-quality examples? Below that, collect or generate more data before training.

Fine-tuning earns its cost when the task is narrow, repeated at high volume, and benefits from shorter prompts: support-agent tone locked to brand, extraction into a fixed schema, classification against a long taxonomy, or distilling a larger model's behavior into a cheaper small one. Our LLM fine-tuning best practices guide for 2026 covers the full production checklist.

How a fine-tuning project runs, step by step

1. Build the eval set first. Hold out 50–200 examples that mirror production traffic, choose a grader (exact match for classification, an LLM-as-judge with a fixed rubric for open-ended tasks), and record a baseline score from the strongest base model with your best prompt. That baseline is the number the fine-tune must beat.
2. Prepare the training data. OpenAI's format is JSONL: one JSON object per line, each holding a chat-style messages array with system, user, and ideal assistant turns. Quality beats quantity — 50 to 500 consistent, well-curated examples is the productive range for most tasks, and a few sloppy examples do disproportionate damage (Respan).
3. Upload and start the job. Upload the file, then create a fine-tuning job naming the base model and method. Start with roughly 2–3 epochs and the default learning-rate and batch-size settings; adjust only if evaluation demands it.
4. Evaluate, then iterate. Score the fine-tuned model against your held-out set. If it does not measurably beat the prompted baseline, you have a more expensive model, not a better one.
5. Deploy with a rollback path. The finished job returns a model ID you call like any other model. Keep the base-model code path live so you can A/B test and revert instantly.

What fine-tuning costs

Costs come in three parts:

• Training. Reinforcement fine-tuning on o4-mini is billed at $100 per training hour, with tokens used for grading charged at the model's per-token rates (OpenAI pricing). Practitioner estimates put a typical supervised job on a mini-class model in the $5–50 range (Respan).
• Inference. Fine-tuned models bill per token like their base models; GPT-4o mini, for reference, lists at $0.15 per million input tokens and $0.60 per million output tokens (OpenAI model docs).
• The hidden line item. Engineering time for data curation and evaluation usually dwarfs the API bill. Budget for it.

Enabling data sharing on a fine-tune job earns inference discounts, at the cost of letting OpenAI train on your data — rarely acceptable in regulated industries.

Where fine-tuned AI agents earn their keep

• Specialized support agents in regulated finance. An agent tuned on compliant response formats and banking terminology answers routine queries in a consistent, auditable tone — the kind of workload our AI agents for financial services page covers in more depth.
• Legal document review. Tuned on statutes, house clause language, and review conventions, an agent summarizes contracts and flags non-standard clauses for human review.
• Marketing and sales voice. Tuned on your best copy and personas, an agent drafts outreach that sounds like your brand rather than like the internet's average — the same specialization logic behind custom GPT agents.
• Reporting agents. Tuned on internal data schemas and report formats, an agent turns natural-language questions into consistently structured reports.

In each case the pattern is identical: a narrow, repeated task, a fixed standard of "good," and enough volume that a smaller tuned model beats a larger prompted one on cost and consistency.

Frequently asked questions

Is OpenAI fine-tuning still available?

Partially. New organizations can no longer create fine-tuning jobs. Existing customers can start new training runs until early January 2027, and already-trained models keep serving inference until their base models are deprecated.

How much training data do I need?

The API accepts as few as ten examples, but real gains start around 50, and 50–500 high-quality, consistent examples cover most tasks. More noisy data hurts; a smaller, ruthlessly curated set wins.

Should I fine-tune or use RAG?

They do different jobs. Fine-tuning teaches behavior — tone, format, task patterns. RAG supplies knowledge — current facts pulled from your documents at request time. Most production agents in knowledge-heavy domains need RAG; fine-tuning is the optional second layer for behavior.

Can I fine-tune GPT-5 or Claude?

No. OpenAI's GPT-5 family is not fine-tunable, and Anthropic does not offer customer fine-tuning for Claude — Claude agents are specialized through prompts, skills, tools, and retrieval instead.

What does a typical project cost?

The training API bill for a supervised job on a mini-class model is usually in the tens of dollars, and inference bills at per-token rates. The real cost is engineering time for data curation and evaluation, so budget people before tokens.

Where to go from here

Fine-tuning is a precision tool: powerful for narrow, high-volume behavior, the wrong answer for knowledge, and now time-boxed on OpenAI's platform. If you are weighing a tuned model against retrieval, tool use, or a managed agent stack, our enterprise AI team helps banks, insurers, and other regulated organizations make that call — from private deployment to AI governance.