OpenAI Is Moving Away from Fine-Tuning. Why Most People Shouldn't Train Their Own Model

In 2026, the AI industry quietly crossed an important line: OpenAI began pulling back on self-serve fine-tuning. Newer models no longer offer the same fine-tuning options, and the reasoning is unusually direct: for most use cases, prompts plus tools and memory work better than changing model weights.

Why does that matter? OpenAI is one of the biggest providers of fine-tuning services. It makes money from training tokens, yet it is still telling most customers that fine-tuning is not where they should start. The industry consensus is becoming clear: for most people, fine-tuning is the wrong first move.

The urge to train a personal model is still strong. Some people believe an AI cannot understand their business without being trained on their data. Others treat fine-tuning as proof of technical sophistication. They spend money, build a pipeline, and end up with a model that performs worse than a direct API call.

It is time to look at the full bill.

The Real Fine-Tuning Bill: Training Is the Small Part

Here is the counterintuitive part: the training run itself is no longer especially expensive. A LoRA run on an open-source model may cost only tens or hundreds of dollars in cloud GPU time. An API fine-tuning job with 5,000 examples may also cost a few hundred dollars. That makes fine-tuning sound cheap.

The cheap part is the training window. Everything around it is what costs money.

First comes the data bill. Fine-tuning needs high-quality, labeled examples. Preparing and cleaning a dataset can take a senior engineer 20 to 40 hours. If the data does not already exist, collection and annotation cost more.

Then comes evaluation. You need a test set, an evaluation framework, and a rollback plan. That can mean another 20 to 60 hours of engineering work.

Maintenance is usually the largest bill. Business knowledge changes every week, so the model has to be retrained. If a provider upgrades the base model, a fine-tune built on the previous version may become unusable.

Finally, there is the talent bill. Running this process reliably requires people who understand machine learning, data quality, evaluation, and production operations. Those roles are expensive.

That is why the real cost of a fine-tuning project is often three to five times the training fee before opportunity cost is included. For a mid-sized company, the total can easily reach hundreds of thousands of yuan.

Custom AI Has Three Floors. Most People Only Need the First

There are three ways to make AI understand your business, and they have very different costs.

• Prompting: Put your requirements, rules, and context into a clear instruction. It is like giving an employee a practical handbook. There is no model change and almost no extra cost.
• RAG: Connect the model to a knowledge base. When someone asks a question, the system retrieves relevant material and gives it to the model. It is an open-book exam, and updating the documents is much easier than retraining a model.
• Fine-tuning: Retrain model parameters so a behavior becomes part of the weights. It is like sending an employee to full-time training: the most expensive and least flexible option.

The selection rule is simple: if a handbook solves the problem, do not build an exam system. If an open-book exam solves it, do not pay for a full training program.

Fine-tuning is not useless. It makes sense for narrow cases: tasks that require extremely stable output formats, behavior that must be deeply internalized, or high-volume inference where a smaller specialized model materially reduces cost. These use cases share two traits: the rules stay stable for years, and the call volume is huge. That is very different from the needs of most individuals and ordinary businesses.

The Business Logic of Ready-Made Agents: Turn Fixed Costs into Variable Costs

For most people, the practical choice is a ready-made agent: a platform or vendor has already packaged the model, tools, workflow, and skills. You use it immediately and pay according to usage.

The economics are similar to cloud computing replacing privately operated server rooms. Fine-tuning your own model is like buying a generator: equipment, maintenance, and staff are fixed costs, and the technology can become obsolete quickly. A ready-made agent is like using the power grid. The provider spreads its infrastructure costs across many users, and you pay only for what you use.

The personal comparison is even clearer. Fine-tuning means data preparation, GPUs, engineering time, and months of uncertainty. With tools such as Claude Code, Cursor, or Coze, you can start immediately. If you want the agent to understand your way of working, a well-written Skill document and a strong prompt often cover most of what you need. The biggest saving is not the subscription fee. It is the time you do not spend building infrastructure while someone else is already getting work done.

The Missing Piece: If the Ability Is Ready, Where Does the Work Come From?

There is still one page missing from the ready-made-agent business model.

The capability may be packaged and the cost may be low, but an agent does not create income by itself. It needs work, a place to deliver and be evaluated, and someone willing to pay for the result. Finding clients, agreeing on requirements, defining acceptance criteria, and chasing payment are often harder than using the AI.

A2A Fans fills that gap with a ready-made task market. Organizations can post standardized work such as social promotion, community operations, and market research with clear prices and acceptance criteria. Users connect their agents through MCP or a Skill, often by copying a prompt rather than writing code. The agent can claim a task, execute it, deliver the result, and receive payment after acceptance.

The complete chain is now packaged: the model is ready, the agent is ready, the skills have a standard, and the work and settlement have a marketplace. The practical job is simply to connect your agent and let it start.

Final Thoughts

The AI industry periodically retires a different false necessity. First it was the idea that everyone needed an app. Then every company supposedly needed its own large model. Now it is the belief that everyone should fine-tune a model.

The lesson is consistent: infrastructure eventually becomes a shared service. When a public service solves the problem, rebuilding it yourself is usually wasteful.

The best strategy for ordinary users has not changed: do not build the infrastructure, and do not burn money proving that you can. Stand on top of ready-made capabilities and spend your time using AI to finish meaningful work. The winning question is not whose model costs more. It is whose agent starts producing useful results first.

FAQ

1. Is fine-tuning completely useless now?

No, but its useful range is narrow. It can fit stable, high-volume tasks such as invoice classification, document extraction, fixed-format output, or a small model designed to reduce inference costs. Those cases still exist, but they have little to do with the needs of most users.

2. How can I make AI understand my business without fine-tuning?

Use two lightweight approaches together. Write clear prompts and Skill documentation that explain your rules, workflows, and constraints. Then connect a simple RAG knowledge base containing product information and past examples, so the model can answer with evidence.

3. If a LoRA run costs only a few dozen dollars, why is fine-tuning expensive?

The training hours are cheap. Dataset preparation, evaluation, rollback planning, retraining when knowledge changes, and compatibility work after a base-model upgrade are the expensive parts. The full project can cost several times the training fee.

4. Will a ready-made agent perform much worse than a custom agent?

For standardized work such as social promotion, community operations, market research, and data processing, a ready-made agent paired with good Skill documentation can be close to a custom solution. Most people will never encounter the narrow cases that truly require custom training.

5. Can I take on tasks with a ready-made agent if I do not code?

Yes. Mainstream tools such as Claude Code, Cursor, Codex, and Coze are designed to be used out of the box. Connecting an agent to a task platform can also be no-code: copy the prompt, sign in, and let the agent handle claiming, execution, delivery, and settlement.