Can You Fine-Tune Claude? What to Do Instead

No. As of 30 October 2026, the Claude API offers no fine-tuning, and Anthropic's own documentation says so. The one route Anthropic documented, fine-tuning Claude 3 Haiku on Amazon Bedrock, has closed: the model reached end of life on Bedrock on 10 September 2026 in the only region that offered its fine-tuning. A team that wants a model trained on its own work trains an open model on its own data and serves it behind an OpenAI-compatible endpoint, so the application changes a base URL and a model name.

This guide covers what Anthropic's documentation says, what happened to the Bedrock route, what Claude fine-tuning costs, where OpenAI's fine-tuning stands, the rule on training with Claude's outputs, and how to train a model on your own data. Every date and quote below was read at the vendor's own documentation on 30 October 2026.

Does the Claude API offer fine-tuning?

No. Anthropic's glossary states that "The Claude API does not currently offer fine-tuning", and points anyone interested to their Anthropic contact. Anthropic's developer documentation describes no fine-tuning endpoint as of 30 October 2026.

The ways to shape Claude's behaviour all work at request time: the system prompt, examples in the prompt, retrieved documents, and prompt caching for long context that repeats. Each one changes what Claude reads on an API call and leaves the model as Anthropic trained it. That context is also billed on every API call, at a discount when cached. For when packing context into Claude is enough and when a trained model wins, see Claude Projects vs a fine-tuned model.

What happened to Claude fine-tuning on Amazon Bedrock?

It closed in 2026. AWS opened fine-tuning for Claude 3 Haiku on Amazon Bedrock in July 2024, as a preview in the US West (Oregon) region. Anthropic's legal summarization guide still reads "Fine-tuning is currently only available through Amazon Bedrock", and Bedrock's fine-tuning page still lists Claude 3 Haiku in Oregon as the only Claude model. The model itself has since been retired.

Bedrock's lifecycle rules say that once a model is in Legacy, "you cannot create new fine-tuning jobs on that model", so new Claude fine-tuning jobs stopped in March 2026. Since 10 September, the Claude 3 Haiku model card says requests in most regions "now fail with ResourceNotFoundException". No newer Claude model has been added to Bedrock's fine-tuning list.

How much does Claude fine-tuning cost?

No route is open, so there is no current price. Bedrock's pricing page lists customization prices for Amazon, Cohere and Meta models and none for Claude as of 30 October 2026, and Anthropic's API has no fine-tuning to price.

The cost worth working out is the alternative's, and it comes in two parts. Training is a one-off cost per model version: preparing examples, running the training, and evaluating the result. Running cost is per task, and a trained model lowers it on two dimensions at once:

1. A lower price per token. A model trained for one job carries only the capability that job needs, so it costs less to serve than a frontier model.
2. Fewer tokens per task. The instructions, examples and context you repeat in every Claude prompt are trained into the model, so they stop being sent, and a model trained on the task answers without hidden reasoning tokens.

How much each dimension saves depends on how much of each prompt repeats and on your volume. Put your own token counts and prices into the cost-per-task calculator, which splits a saving into its price effect and its volume effect. For the margin arithmetic at product scale, see AI startup gross margins and inference costs.

What does Anthropic's policy say about training on Claude's outputs?

It prohibits it without Anthropic's prior authorization. Anthropic's Usage Policy, in the version effective 12 November 2026, prohibits using its products to "Use outputs to train an AI model (e.g., 'model scraping' or 'model distillation') without prior authorization from Anthropic". The rule covers any model you would train, whether or not it competes with Claude.

The data that gives you a model worth owning is your own anyway: your records, the answers your team accepted, the corrections they made. A model distilled from a public API can be rebuilt by anyone with the same API key, while one trained on your work stays yours. Our post on model distillation vs fine-tuning covers the providers' terms and why your own data builds the moat. This is a general overview and is not legal advice.

Can you fine-tune ChatGPT?

Only through the OpenAI API, and that route is closing too. Fine-tuning OpenAI's models happens on the OpenAI API's fine-tuning platform, and OpenAI's supervised fine-tuning guide now opens with "OpenAI is winding down the fine-tuning platform."

How can I fine-tune OpenAI models?

Only as an eligible existing customer, and only until 6 January 2027. In OpenAI's words: "The platform is no longer accessible to new users. Eligible existing customers can create training jobs until January 6, 2027, unless the base model shuts down earlier." On 14 January 2027, GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down, and per OpenAI's deprecations page, "Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models."

OpenAI's open-weight gpt-oss models sit outside that schedule: you download the weights and train them like any other open model. If you have fine-tunes on OpenAI today, our OpenAI fine-tuning migration guide covers the deadlines and a four-step move.

How can I fine-tune an AI model on my own data?

Train an open model on examples from your own work, evaluate it against the model you use today, and serve it behind an OpenAI-compatible API. Open models have no vendor-set retirement date, so the trained version runs for as long as you need it.

1. Pick one task. Choose a task that runs at volume and whose output your team can check: a classification, an extraction, a summary in a fixed format, a reply in your house style.
2. Collect examples from your own work. Pair real inputs with the outputs your team accepted. Hold some back as an evaluation set the model never trains on.
3. Train an open model. Choose the smallest open model that holds your quality bar on the held-out set, and check its licence allows commercial use.
4. Compare it with Claude on your evaluation set. Score both on the same held-out examples, with the measure your team already uses to judge the work.
5. Serve it behind an OpenAI-compatible API. Your application changes the base URL and the model name, and the rest of the code stays as it is.

Training on your own task can close the gap to a frontier model on that task. In our QMSum study (arXiv 2609.25028), trained open models scored above five hosted proprietary models on ROUGE-1 for query-focused meeting summarization, three of them Claude models (Claude Opus 5, Claude Sonnet 5 and Claude Haiku 4.5). The hosted models ran zero-shot under one concise prompt and wrote longer answers, which contributes to the gap on a reference-overlap metric, and no human or factuality evaluation was run. For how open and frontier models compare in general, see open-weight vs frontier models.

Ertas runs this as the adaptive model endpoint: a model trained on your task and your team's feedback, served through an OpenAI-compatible API. Frontier models answer at the start, the trained model takes over once it passes your evaluations, and hard cases stay on frontier models. New versions train on your agents' traffic in the background, are scored against your evaluations, and go live only with your sign-off; if a new version falls short, the current model stays in service. It runs Ertas-managed, in your own cloud or on-premise. The endpoint is currently invitation only, and we are developing a more streamlined version for wider access.

Frequently Asked Questions

Is there a Claude fine-tuning API?

No. As of 30 October 2026, Anthropic's documentation states that the Claude API does not currently offer fine-tuning, and suggests asking your Anthropic contact if you want to explore it.

Can I still fine-tune Claude 3 Haiku on Amazon Bedrock?

No. Claude 3 Haiku entered Legacy on Bedrock on 10 March 2026, which stopped new fine-tuning jobs, and reached end of life in Oregon, its only fine-tuning region, on 10 September 2026.

Can I fine-tune Claude Sonnet or Claude Opus?

There is no documented route as of 30 October 2026. Neither the Claude API nor Bedrock's fine-tuning list offers fine-tuning for any current Claude model.

Is it allowed to train a model on Claude's outputs?

Only with Anthropic's prior authorization. The Usage Policy effective 12 November 2026 prohibits using outputs to train an AI model, naming "model scraping" and "model distillation" as examples. Train on your own data instead.

What is the best alternative to fine-tuning Claude?

An open model trained on your own examples for one task, evaluated against Claude on a held-out set, and served behind an OpenAI-compatible API. It lowers cost per task on both dimensions, price per token and tokens per task, once the task runs at volume.

The Short Version

You can't fine-tune Claude as of 30 October 2026. The Claude API has no fine-tuning, and the Bedrock route for Claude 3 Haiku closed when that model was retired in 2026. Training another model on Claude's outputs needs Anthropic's prior authorization. The route that works is an open model trained on your own data, evaluated against Claude on your task, and served behind an OpenAI-compatible endpoint. OpenAI's fine-tuning platform closes to new training jobs on 6 January 2027, so the same route applies there.

---

Sources, read 30 October 2026: Anthropic, Glossary, Legal summarization guide, Model deprecations and Usage Policy (effective 12 November 2026); AWS, Fine-tune Anthropic's Claude 3 Haiku in Amazon Bedrock (10 July 2024), Bedrock fine-tuning models, model lifecycle, Claude 3 Haiku model card and pricing; OpenAI, supervised fine-tuning guide and deprecations. Our research: the QMSum study.