Fine-tuning used to sound like the serious, grown-up way to customize AI. In 2026 it is the option with the shortest shelf life, and most business problems never needed it.
June 30: Anthropic releases Claude Sonnet 5. September 22: OpenAI releases GPT-6 Sol and GPT-6 Luna. September 28: Anthropic releases Claude Sonnet 5.5, roughly three months after its predecessor. October 23: OpenAI shuts down a batch of older models, including fine-tuned versions built on GPT-4 and GPT-3.5 Turbo. January 6, 2027: OpenAI stops letting even its active customers start new fine-tuning jobs.
That is one quarter of the AI market. Read it as a planning document and one conclusion jumps out. Anything you train on top of a specific model is tied to a product with a lifespan measured in months.
Three ways to customize, in plain terms
When a business wants AI to behave like it knows the company, there are three basic tools.
Prompting means giving the model instructions and examples inside the request. Your tone guide, three sample answers, the rules for what it should never say. Nothing about the model changes. You change what you hand it.
Retrieval means the system looks up your information at the moment it answers. Product specs, policies, past tickets, contracts. The model reads the relevant pages and responds using them. Update the source document and tomorrow’s answers change.
Fine-tuning means training the model on hundreds or thousands of your own examples so its internal behavior shifts. The customization lives inside a custom copy of the model.
All three are legitimate. They solve different problems, and they age very differently.
The fine-tuning market just moved
OpenAI’s own deprecation notices tell the story. Since May 7, 2026, organizations that had never fine-tuned a model can no longer start. Since July 2, organizations that have not used a fine-tuned model in the past 60 days are locked out too. On January 6, 2027, new fine-tuning jobs end for everyone on OpenAI’s self-serve platform. Existing fine-tuned models keep running only until their underlying base model is retired.
Anthropic has never offered fine-tuning in the Claude API itself. Its fine-tuning path has run through Amazon Bedrock, historically on Claude 3 Haiku, a small model several generations behind Sonnet 5.5 and Opus 5.5.
Google continues to offer supervised fine-tuning for Gemini models through Google Cloud. Open-weight families such as Meta’s Llama, Mistral, and Alibaba’s Qwen can be fine-tuned by anyone with the hardware or a hosting partner, and you own the result.
So the vendors behind the two most widely deployed model families, OpenAI and Anthropic, are steering customers away from training custom copies and toward prompting and retrieval on their newest models. That is a signal worth taking seriously, whatever you think of their reasons.
A fine-tune is a depreciating asset
Here is the part that rarely makes it into the pitch deck. A fine-tuned model is frozen on the base model it was trained from. When the vendor ships a better base model, your custom version does not upgrade. It stays where it is until the vendor retires it.
Anthropic shipped Sonnet 5 and Sonnet 5.5 about three months apart. A team that spent July and August training a custom version of a mid-tier model would have finished just in time to watch a faster, cheaper base model ship without their customizations.
Prompts and retrieval move with you. A good prompt written for Sonnet 5 usually works on Sonnet 5.5 with light adjustment. A retrieval system pointed at your documents can swap the model underneath in an afternoon. A fine-tune has to be rebuilt from scratch.
Facts versus behavior
The most common reason companies ask about fine-tuning is to make the AI know their business: products, pricing, procedures. That is the job fine-tuning does worst.
Fine-tuning teaches behavior. Format, tone, a narrow classification task, a strict output structure. It is a poor way to store facts, because facts change and the training does not. When your price list updates in January, a fine-tuned model keeps quoting December until someone retrains it.
Retrieval handles facts. Prompting handles most behavior. Fine-tuning earns its place only for behavior that prompting cannot hold reliably, at a volume where the cost of training pays back.
The industry already voted on this. The Menlo Ventures 2024 enterprise survey found 51% of enterprise AI deployments using retrieval in production, against 9% relying primarily on fine-tuning. That gap has only widened as context windows grew and newer models got better at following written instructions.
The order of operations
Start with prompting. Write the instructions, include real examples, test against a set of actual cases from your business. A surprising share of projects stop here and work fine.
Add retrieval when the answers depend on information that changes or that is too large to paste in. This is where most serious business systems end up.
Consider fine-tuning last, and only when four things are true. The task is narrow. The behavior you need is stable. The volume is high enough that a smaller tuned model would save real money over a larger general one. And prompting plus retrieval has already been tested and measured against the same cases, so you know exactly what gap training would close.
When those conditions hold, an open-weight model you control often makes more sense than a vendor fine-tune, because you decide when it gets retired.
Questions to ask anyone pitching a fine-tune
When a vendor or an internal team proposes a custom model project, five questions sort the real cases from the expensive ones.
What did prompting alone score on our actual cases? A proposal without that number has skipped the cheapest test.
Which failures are about facts, and which are about behavior? Fact failures belong to retrieval. Only behavior failures are candidates for training.
Which base model are we training on, and what is its retirement date? Get the vendor’s deprecation policy in writing before the first training run.
What happens to our training data and our tuned model when that base model is retired? Somebody has to own the rebuild, and it should be budgeted now.
Can we run the same test set against the next model release in a day? A team that cannot answer yes is building something that will be hard to maintain.
None of these questions are hostile. A good partner answers all five without hesitation, and some of the answers will point toward fine-tuning. Most will point somewhere simpler.
What this looks like in Wisconsin
Wisconsin businesses are behind on AI adoption, and that creates pressure to catch up fast. The Census Bureau’s Business Trends and Outlook Survey, in its round covering late August into early September, found 18.3% of Wisconsin businesses had used AI in the prior two weeks, 41st among the states. Nationally the figure was 23.8%, up from 17.7% at the start of the year. Wisconsin moved only about one point over the same stretch.
At the same time, BBB Wisconsin’s 2026 business survey found half of respondents believe AI can make their business more competitive.
Put those together and you get a familiar pattern: a company that feels late, decides to make up ground with something ambitious, and signs up for a custom model project. That is exactly the instinct to resist. The fastest way to catch up is the cheapest, most portable path. Prompting and retrieval get a Wisconsin manufacturer or distributor to a working system in weeks, on whichever model is best this quarter, with no training run to throw away when the next release lands.
The honest tradeoff
Fine-tuning still has real uses. A high-volume classification task processing tens of thousands of documents a day can justify a small tuned model on cost alone. A strict output format feeding downstream software sometimes needs it. Those cases exist, and we build them.
But most requests we hear for fine-tuning are really requests for an AI that knows the business. That is a retrieval problem with a prompting problem on top.
Before you fund a training project, ask for the version without one. Build the prompt and retrieval system first, measure it on your real cases, and make the fine-tuning decision with numbers in hand.
Planning a 2027 AI budget and getting pitched custom model work? Tell us what problem you are trying to solve, and we will tell you the simplest thing that solves it.
