I Checked the Anthropic Usage Policy Update (Effective November 12, 2026) That Arrived via AWS Health
This page has been translated by machine translation. View original
Introduction
An AWS Health notification dated October 14, 2026 announced an update to the Usage Policy for Anthropic models available on Amazon Bedrock. The effective date is November 12, 2026.
This article uses the AWS Health API to check the notification type and body, and summarizes an impact checklist by use case along with the main changes between the old and new policies.
The notification that arrived in Health
The AWS Health API is available with any of the following support plans: AWS Business Support+, Enterprise Support, or Unified Operations. For accounts that have not yet migrated to the new support plans, the legacy Business, Enterprise On-Ramp, and Enterprise Support plans can also be used. Accounts without one of these plans receive SubscriptionRequiredException. In that case, check the event in the AWS Health Dashboard.
Using describe-events in the AWS Health API, I retrieved events filtered to the BEDROCK service.
aws health describe-events --region us-east-1 --filter services=BEDROCK
Here are the fields from the retrieved event that indicate the nature of the notification.
{
"service": "BEDROCK",
"eventTypeCode": "AWS_BEDROCK_OPERATIONAL_NOTIFICATION",
"eventTypeCategory": "accountNotification",
"startTime": "2026-10-14T15:20:00+09:00",
"statusCode": "open",
"eventScopeCode": "ACCOUNT_SPECIFIC",
"actionability": "ACTION_REQUIRED"
}
Because eventScopeCode is ACCOUNT_SPECIFIC, this is not an announcement for all accounts but a notification in which your own account has been identified as affected. actionability is ACTION_REQUIRED, but the response requested is to review the policy, not to perform any operation in the AWS console.
I retrieved the body with describe-event-details. For <EVENT_ARN>, specify the event ARN included in the describe-events result.
aws health describe-event-details --region us-east-1 \
--event-arns <EVENT_ARN> \
--query 'successfulSet[0].eventDescription.latestDescription' \
--output text
Here is the part of the body about the policy update.
Effective November 12, 2026, Anthropic is updating its Usage Policy [2]. The updated policy governs how Anthropic models — including the Claude model family — may be used, and applies to all Anthropic models made available through Amazon Bedrock. We recommend reviewing the updated policy before the effective date to ensure your applications and end-user workflows remain compliant.
What you should do?
- Review the updated Anthropic Usage Policy [2] and compare it against your current use cases.
- Share the updated policy with internal stakeholders (legal, security, product) who govern your use of generative AI.
- Continued use of Anthropic models on Amazon Bedrock on or after November 12, 2026 constitutes acceptance of the updated terms.
There is no change to model availability, pricing, API behavior, or model IDs as a result of this update. No action is required in the AWS console.
[2] https://www.anthropic.com/legal/aup
What is requested is to review the updated policy against your current use cases and share it with stakeholders in legal, security, and product. If you continue to use Anthropic models on Bedrock on or after November 12, 2026, you are deemed to have accepted the updated terms.
Note that this update does not change model availability, pricing, API behavior, or model IDs. No action is required in the AWS console either.
The old and new policies I checked
The new version is the one effective November 12, 2026. I checked the version published at /legal/aup as of October 16, 2026.
The old version is the one effective September 15, 2025. The archive page I checked on October 16, 2026 shows the notice "This document has been replaced by a newer version", indicating it is the old version.
Impact check by use case
Here are the use cases that deserve particular attention following this update. If one applies to you, do not decide based only on the summary below; check the relevant section of the official policy.
| Use case | Key point of the update | What to check |
|---|---|---|
| Internal drafting, research, summarization, and analysis | Explicitly stated as excluded from the High-risk requirements | Whether that output is ultimately provided as advice or decisions for individuals |
| Advice and decisions for individuals | Covered areas reorganized (from 7 areas to 11 areas) | Check the official policy to see whether your use case applies |
| Integrations that operate equipment without human approval | High-risk Physical Actions newly established | Check the applicability conditions and safety requirements in the official policy |
| Resale and proxying of Claude | Provision through unauthorized means prohibited | Check whether your form of service falls under this |
For whether something applies and for the specific requirements, check the original text of the Anthropic Usage Policy.
Main changes between the old and new versions
Advice and decisions for individuals (High-risk AI Recommendations)
In the new version, model output used for advice to individuals or decisions about individuals is organized as High-risk AI Recommendations, and the covered areas went from 7 to 11. This is a reorganization that includes splitting and replacing areas.
In addition, internal drafting, research, summarization, and analysis, as well as business operations that do not involve advice or decisions about specific individuals, are now explicitly stated as excluded from High-risk AI Recommendations. However, even if the output is for internal use, you need to check applicability if it is ultimately provided as advice or decisions for individuals.
For the covered areas and exclusions, as well as specific requirements such as review by qualified professionals and disclosure of AI use, check the High-risk AI Recommendations section of the official policy. Note that the disclosure must clearly convey that AI was used, but it does not need to name Anthropic or Claude.
Control of equipment and robots (High-risk Physical Actions)
The new version introduces High-risk Physical Actions. It covers uses where model output controls hardware without human approval, where that equipment can perform operations that may harm people. If it applies, requirements include a setup in which a qualified person can stop it at any time, and enforcing operational limits in equipment or a controller independent of the model's output. Uses in which a qualified person reviews model-generated code or action plans before execution are out of scope.
For the types of covered equipment, exemptions, and specific safety requirements, check the High-risk Physical Actions section of the official policy. You must also separately satisfy legal and certification requirements related to vehicles, aviation, medical devices, machinery, and occupational safety.
Resale and proxying of Claude
A new prohibition has been added to the Do Not Abuse Our Platform section of the new version. It covers reselling, proxying, or providing access to Claude through unauthorized means. Examples given include services that relay requests via consumer subscriptions and services that disguise the product or client being used. If you operate a service that provides third parties with access to Claude, check the relevant part of the official policy.
Decisions on law enforcement and criminal justice
The scope of prohibition on decisions related to law enforcement and criminal justice has been expanded. If your use case falls under this, check the Universal Usage Standards section of the official policy.
Summary
No configuration changes in the AWS console are needed because of this update.
The main changes are as follows.
- Revision of High-risk requirements for advice and decisions for individuals (internal drafting, research, summarization, and analysis that is not ultimately provided as advice or decisions for individuals is explicitly stated as excluded)
- New requirements for equipment control without human approval
- Prohibition of reselling and proxying Claude through unauthorized means
Please review the updated Usage Policy before the effective date. If your use may fall under the prohibited uses or the scope of the High-risk requirements, we recommend consulting relevant departments such as legal, security, and product to avoid usage that violates the policy.
