AWS Bedrock explains model lifecycle: Active, Legacy and End-of-Life phases

On April 9, AWS published an official guide to the lifecycle of foundation models in Amazon Bedrock — a response to the question customers keep asking: “What if AWS retires a model my production relies on?” The answer is now structured around three clearly defined phases.

The three model states

ACTIVE is the standard state — the model receives maintenance and updates, and is available to all users without restrictions.

LEGACY is the deprecation phase, which lasts 6 months before final retirement. Users are notified by email, through the AWS Health Dashboard, Bedrock console alerts and programmatically via the API. For models with an EOL date after February 1, 2026, AWS is adding a new concept — an extended access period that lasts at least an additional 3 months. During this period, active users continue to have access, but quota increases via AWS Service Quotas will likely not be approved. Pricing may change with advance notice.

END-OF-LIFE means complete retirement — the model is no longer available, not for calls, not for fine-tuning, and not for any existing workload.

A 5-step migration playbook

AWS recommends a structured approach to migration: (1) assessment of current usage — which applications call which model, and how often; (2) research of replacement models — compare capabilities and pricing; (3) testing — parallel runs, shadow testing, A/B testing; (4) gradual implementation — phased rollout instead of a big-bang switch; (5) post-migration monitoring — watch for regressions after the switch.

Technically, migration usually involves updating API references, requesting quota increases for the new model, adjusting prompts (since different models react differently to the same prompt), updating response parsing logic, and optimizing token consumption.

What to check

The most important practical point AWS emphasizes: make sure account contact information and notification settings reach the right people. The worst-case scenario is a deprecation notice ending up in the inbox of someone who no longer works for the organization, resulting in the model being shut off without warning to the production team.