AWS Security ChangesHomeSearch

AWS bedrock: Add preserved thinking binding and fallback credit beta to Sonnet 5.5 card

Service: bedrock · 2026-11-06 · Documentation medium

File: bedrock/latest/userguide/model-card-anthropic-claude-sonnet-5-5.md · Type: authz

Summary

Updates the Reasoning bullet (noting thinking.type 'disabled' is unsupported and 'between_tools' should be used) with a link to Adaptive thinking, adds a 'Preserved thinking' bullet stating thinking blocks are bound to the model, conversation prefix, and the producing AWS account, and adds a beta section describing fallback credit tokens for refused requests with supported endpoints and valid fallback models.

Security assessment

The preserved-thinking note documents account-level binding of reasoning blocks (cross-account replay protection) and the fallback-credit beta describes billing behavior for retried refused requests; both are security-adjacent documentation of isolation and request-handling behavior, with no reference to a specific vulnerability or CVE.

Evidence

+  * **Preserved thinking:** Thinking blocks are bound to the model and the conversation prefix that produced them. Replay earlier turns unchanged, or a request returns a 400 error. On Claude Sonnet 5.5, thinking blocks are also bound to the AWS account that produced them. For more information, see [Preserved thinking](claude-messages-thinking-block-binding.html).

Diff

diff --git a/bedrock/latest/userguide/model-card-anthropic-claude-sonnet-5-5.md b/bedrock/latest/userguide/model-card-anthropic-claude-sonnet-5-5.md
index b27960c..7b2699a 100644
--- a//bedrock/latest/userguide/model-card-anthropic-claude-sonnet-5-5.md
+++ b//bedrock/latest/userguide/model-card-anthropic-claude-sonnet-5-5.md
@@ -35 +35,3 @@ Claude Sonnet 5.5 is a smarter, more efficient Sonnet. It's a clear leap forward
-  * **Reasoning:** Supported (adaptive thinking is on by default; effort level configurable — low, medium, high, xhigh, max; default: high)
+  * **Reasoning:** Supported (adaptive thinking is on by default; `thinking.type: "disabled"` is not supported, so use `"between_tools"` to turn off up-front thinking at `high` effort or below; effort level configurable — low, medium, high, xhigh, max; default: high) See [Adaptive thinking](claude-messages-adaptive-thinking.html).
+
+  * **Preserved thinking:** Thinking blocks are bound to the model and the conversation prefix that produced them. Replay earlier turns unchanged, or a request returns a 400 error. On Claude Sonnet 5.5, thinking blocks are also bound to the AWS account that produced them. For more information, see [Preserved thinking](claude-messages-thinking-block-binding.html).
@@ -176,0 +179,8 @@ Yes | `compact-2026-09-04`
+**Fallback credit for refused requests (beta)**
+
+When Claude Sonnet 5.5 refuses a request, the response can include a fallback credit token. Redeem the token when you retry the same request on a fallback model, and the retry is billed for the cached conversation prefix at cache-read rates instead of cache-write rates. For more information, see [Fallback credit for refused requests (beta)](claude-messages-fallback-credit.html).
+
+**Supported endpoints** | **Beta value** | **Valid fallback models**  
+---|---|---  
+`bedrock-runtime`, `bedrock-mantle` | `fallback-credit-2026-07-01` | Claude Sonnet 5 (`anthropic.claude-sonnet-5`)  
+