For the complete documentation index, see llms.txt. This page is also available as Markdown.

Auto Update [BETA]

We organize our models into model series. Each series groups related model versions together under a shared name. When we release a new version within a series, we add it to that series on our platform.

By using the :latest tag on a series name, your application always points to the most recent model in that series no manual updates required.

How It Works

Every model series has a :latest alias that always resolves to the newest available model in that series. When you use :latest in your requests, we automatically route you to the current latest model. As soon as we publish a new model to a series, any request using :latest will use that new model.

Instead of specifying a fixed model version, append :latest to the series name:

{
  "model": "claude-sonnet-4:latest",
  "messages": [...]
}

That's all you need to do. From this point on, your application will automatically use the newest model in the claude-sonnet-4 series whenever we release one.

Pinning to a Specific Version

If you need predictable, stable behavior you still can pin to a specific model version:

{
  "model": "claude-sonnet-4-5",
  "messages": [...]
}

This version will remain unchanged until you explicitly update it.

Choosing the Right Approach

Approach
Best for

:latest

Staying current automatically, prototyping, non-critical workloads

Pinned version

Production systems, regulated environments, reproducibility requirements

Things to Keep in Mind

Behavior may change between versions. New models in a series are generally improved, but responses may differ from a prior version. If your application depends on highly consistent output, consider pinning to a specific version and testing new models before adopting them.

Updates happen immediately. There is no gradual rollout when using :latest — once a new model is published to a series, all requests using :latest will use it.

Last updated