Amazon Bedrock model retirements this fall: the dates, the notice rules, and what to check

Several foundation models on Amazon Bedrock reach retirement dates this fall. The AWS model lifecycle table gives October 14, 2026 as the end of life (EOL) date for Anthropic's Claude Sonnet 4 in the Regions it lists, which include us-east-1, us-east-2, us-west-1, and us-west-2. Claude Sonnet 4.5 entered the Legacy state on October 8, 2026, with an EOL date of April 8, 2027. AWS also now documents a separate lifecycle policy for models launched on or after September 7, 2026.

A foundation model is the pre-trained model an application calls through the Bedrock API. On Bedrock, every model is in one of three states: Active, Legacy, or End-of-Life. AWS says that on, or soon after, the EOL date, requests to the model fail, and it is direct about who has to act: "Migration will not happen automatically."

The dates

These dates come from the AWS table for models launched before September 7, 2026. Each applies only to the Regions listed in that table for the model.

Claude Sonnet 4: EOL on October 14, 2026.
AI21 Labs Jamba 1.5 Large and Jamba 1.5 Mini: EOL on November 26, 2026.
TwelveLabs Marengo Embed v2.7: EOL on November 30, 2026.
Claude Opus 4.1: EOL on January 8, 2027. Its public extended access period started on October 8, 2026.
Claude Sonnet 4.5: Legacy since October 8, 2026. Public extended access starts on January 8, 2027, and EOL is April 8, 2027.
Already at EOL: Cohere Command R and Command R+ (August 19, 2026), Claude 3 Haiku (September 10, 2026, including us-gov-east-1 and us-gov-west-1), amazon.nova-sonic-v1:0 (September 14, 2026), and amazon.nova-canvas-v1:0, amazon.nova-reel-v1:0, and amazon.nova-reel-v1:1 (September 30, 2026).

Lifecycle state can differ by Region. AWS notes that a model can remain Active in Regions that a row does not list, and that a Region's absence from a row does not mean the model will stay available there indefinitely.

What the Legacy state restricts

New customers cannot use a Legacy model, and existing customers may lose access after 15 days of inactivity. Check any workload that calls a Legacy model only occasionally.
You cannot create new Provisioned Throughput for a Legacy model, or new fine-tuning jobs on it.
For models with EOL dates after February 1, 2026, a model enters public extended access after at least three months in Legacy. AWS says to expect higher pricing in that period, set by the model provider.
After EOL, the only exception AWS describes is a private arrangement between you and the provider for continued access.

What is different for newer models

Under the older policy, a model remains on Bedrock for at least 12 months after launch and spends at least 6 months in Legacy before EOL. For models launched on or after September 7, 2026, the model card sets the terms. Where applicable, the card shows an "EOL no sooner than" date and a Legacy period, which is the notice period before EOL. AWS says there are two Legacy periods, 6 months and 45 days, and that most models have 6 months.

Read the card before you build on a model. A 45-day notice leaves little time to test a replacement. AWS also states that Bedrock lifecycle dates may differ from the dates a model provider publishes.

What to do now

Measure Legacy usage. Amazon CloudWatch has a LegacyModelInvocations metric for Bedrock, described as the number of invocations using Legacy models. Bedrock metrics carry a ModelId dimension, so you can see which model is involved.
Check each Region you use. Call ListFoundationModels or GetFoundationModel in that Region and read the modelLifecycle field. It reports a status of ACTIVE or LEGACY, and has fields for the legacy, public extended access, and end of life times.
Search code, configuration, and infrastructure templates for retiring model IDs. For Claude Sonnet 4 the ID is anthropic.claude-sonnet-4-20250514-v1:0. Its inference profile IDs, such as us.anthropic.claude-sonnet-4-20250514-v1:0, contain the same string.
Choose an Active replacement and test it. The lifecycle table does not name one. Run your own evaluation set, and recheck prompts and cost before you switch.
Keep model IDs in configuration instead of scattered through code, and record each model's EOL date on your roadmap when you adopt it.