Skip to main content

Model deprecation policy

When Crusoe deprecates a model offered through Serverless Inference, you'll receive advance notice and your traffic will automatically route to a replacement model—if one exists. This prevents interruptions to your service and ensures that your code keeps working.

Notice period and notification​

Crusoe gives you at least fifteen (15) days' notice before deprecating a model offered through Serverless Inference. If your project called the model within the last 30 days, you'll receive notifications through the following channels:

  • Email: Sent to project owners and members.
  • Admin Console: A notification appears in the console.
  • Slack/webhook: Sent via configured webhook channels.

You can also check a model's deprecation status with the /v1/models API. Each model's entry includes a shutdown_date field set to the UTC cutover date, or null if no deprecation is scheduled.

Named replacement and cutover details​

Deprecation notices include:

  • The name of the suggested replacement model—if one is available.
  • The cutover date. Cutover happens at midnight (00:00) UTC on that date.

Automatic request routing​

At cutover, Crusoe transitions traffic from the deprecated model to its replacement with automatic request routing. Routed requests are billed at the replacement model's price, not the deprecated model's price. See the pricing page for current pricing.

Response transparency

API responses report the serving model so you always know which model is handling your request. For example, a request to deepseek-v4-flash that routes to deepseek-v4.1-flash will return "model": "deepseek-v4.1-flash".

In-flight requests​

Streaming requests that are already in flight at cutover time finish on their original model. Automatic routing applies only to requests received after cutover.

Handle unsupported model parameters​

If a request uses a feature or parameter that the replacement model doesn't support, the API returns an error message instead of silently changing behavior. Review the replacement model's documentation to identify any breaking differences before migration.

Migration resources​

View the list of recommended replacements for deprecated models.

FAQs​

Short, self-contained answers to the questions users most often ask about Crusoe's Serverless Inference model deprecation process.

How do I check whether a Serverless Inference model is deprecated or scheduled for deprecation?​

You can check the deprecation status of a model through multiple channels:

  • Open the model card in the Intelligence Foundry console to see if a deprecation field displays the cutover date.

  • Check Crusoe Updates for announcements that include affected models, cutover dates, and recommended replacements.

  • If your project has called the model in the last thirty (30) days, you'll also receive direct notifications via email, Admin Console, and configured Slack or webhook channels.

  • For programmatic checks, call GET /v1/models to view the model's entry. The entry carries a shutdown_date field set to the UTC cutover date or returns null if the model isn't set for deprecation. On the cutover date, traffic to the deprecated model switches to its replacement and the deprecated model is removed from the list, so /v1/models only returns models you can call directly.

    Example entry during the notice period:

    {
      "id": "deepseek-v4-flash",
      "object": "model",
      "created": 1710000000,
      "owned_by": "deepseek",
      "shutdown_date": "2026-12-01"
    }

What happens when a replacement model is also later deprecated?​

Routing aliases chain across successive deprecations so your code keeps working through multiple model generations. For example, if deepseek-v4-flash is deprecated in favor of deepseek-v4.1-flash, requests to deepseek-v4-flash automatically route to deepseek-v4.1-flash. If deepseek-v4.1-flash is later deprecated in favor of deepseek-v5, both deepseek-v4-flash and deepseek-v4.1-flash route to deepseek-v5. See Automatic request routing for the full cutover behavior.

How long are requests automatically routed to the replacement model and billed at the replacement's pricing?​

Automatic routing lasts for 90 days. This 90-day window gives you time to update your code to call the replacement model directly without disrupting production traffic. After the 90-day period, Crusoe may remove the routing without further notice.

Policy details​

This policy applies to Serverless Inference and is part of the Documentation for the Crusoe Intelligence Foundry and is subject to the Agreement. Crusoe may update this policy occasionally. Crusoe may deprecate a model on shorter or no notice where required by law, to address a material security risk, or where Crusoe's rights to offer the model are withdrawn or materially changed, and will notify you as soon as reasonably practicable.