The question this series deferred

Three courses have raised fine-tuning and set it aside.

The RAG course said fine-tuning teaches behavior while retrieval supplies facts. The embeddings course said to rule out chunking, hybrid search and reranking first. The few-shot course said examples shape behavior at a fraction of the cost, and to expect not to need fine-tuning.

Every one of those deferrals was correct. Together they lead you to the real question: when is training a model the right call, and what does it cost?

What fine-tuning is

Everything you have built so far changes the input. Retrieval adds facts to the prompt. Examples demonstrate a pattern. Schemas constrain the output. The model is a fixed function you keep feeding better arguments.

Fine-tuning changes the function. You continue training an existing model on your examples, and it comes back with different weights: a model that behaves differently before you send it anything.

flowchart TB
  subgraph Input["Changing the input"]
    P["Prompt + examples + retrieved context"] --> M1["Model, unchanged"]
  end
  subgraph Weights["Changing the model"]
    D["Your dataset"] --> T["Training run"]
    T --> M2["New weights"]
  end

The distinction matters because it decides what you own. Change the input and you own a prompt. Change the weights and you own a model: a training pipeline, a dataset, an evaluation suite, a serving decision, and a version to keep current.

The availability reality, as of October 2026

Before any of the theory, one fact decides whether this is a decision you can make at all.

Anthropic does not expose fine-tuning through its public API. There is no training endpoint alongside Messages, Batches and Files. For a while a Claude model could be fine-tuned through Amazon Bedrock, but only Claude 3 Haiku, and that route has closed: Bedrock stopped taking new customers for that model in March 2026 and retired it on September 10, 2026. Anthropic's own docs say the API does not offer fine-tuning, and to ask your Anthropic contact if you want to explore it.

For open-weight models (the Llama, Mistral and Qwen families and their relatives), fine-tuning is fully available, because you have the weights and can train and serve them yourself.

That splits the practical question in two:

Verify this before planning anything. Availability is the fastest-moving fact in this course (the Bedrock route closed in September 2026), and a plan built on a capability nobody offers you is not a plan.

Why the theory matters anyway

Even where fine-tuning is unavailable, the reasoning is worth having, for three reasons that keep coming back.

You will be asked

"Should we fine-tune a model on our data?" comes up in most teams that touch this technology. The useful answer is a decision procedure.

The concepts transfer

Datasets, splits, overfitting, evaluation against a baseline, and the leakage that invalidates a result work the same whether you are training a model or measuring a prompt. The embedding-model adaptation in the embeddings course was this discussion in miniature.

The alternatives are the real answer

Knowing what fine-tuning would buy is how you recognize that a system prompt, examples and prompt caching already buy most of it. That is what the three deferrals were saying, and lesson 2 makes it precise.

What this course covers

1. The question this series deferred (you are here).
2. What fine-tuning is good at, and what it is not.
3. The dataset is the project.
4. How training works, conceptually.
5. Evaluating a fine-tune honestly.
6. The costs nobody quotes.
7. Putting it together: a decision procedure.

There is no fine-tuning code in this course, on purpose. The API you would call depends on a platform whose details will have moved by the time you need them. The decision, the dataset and the evaluation will still hold.

Key takeaways

• Everything before this course changed the model's input; fine-tuning changes its weights, and therefore what you own.
• As of October 2026 the Anthropic API exposes no fine-tuning endpoint, and Bedrock's Claude 3 Haiku fine-tuning ended when that model was retired on September 10, 2026; anything further means asking Anthropic directly.
• Fine-tuning open-weight models is fully available, with the operations as your responsibility.
• Verify availability before planning; it is the fastest-moving fact in this course.
• The reasoning transfers regardless, because datasets, splits, overfitting and honest baselines apply to prompts as much as to weights.