npm / @ai-sdk/amazon-bedrock
@ai-sdk/amazon-bedrock
What changed in @ai-sdk/amazon-bedrock 3 to 5, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.
4 breaking1 security12 new APIs and options10 behavior and defaults29 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli init@ai-sdk/amazon-bedrock 538 changes
Breaking changes 1
The thinking block binding field was renamed for Bedrock.
Security 1
Sanitizes document filenames before sending them to Bedrock.
New APIs and options 9
Adds Claude Haiku 5.5 support on Amazon Bedrock, with a 128k max output token limit, adaptive thinking, and all five effort levels.
Adds the US-prefixed Claude Haiku 5.5 model ID for Amazon Bedrock.
Portable reasoning now supports max, with native provider mappings, compatibility coercions, and budget-based fallback support.
Adds per-request metadata to the Converse and ConverseStream APIs for tagging Amazon Bedrock invocation logs for cost attribution.
Adds the 20260318 web search and web fetch tools with response inclusion controls.
Claude Fable 5.1 model IDs were added to the Bedrock provider.
Adds a provider option to choose between native structured output and the JSON tool fallback for Anthropic models.
Anthropic reasoning budgets are enabled for application inference profile ARNs.
A model family setting is available for embeddings to support ARN models.
Behavior and default changes 5
Bedrock uses native structured output and strict tools only for known supported Claude models; future Claude model IDs omit unsupported fields and use JSON instructions for structured output by default.
Top-level embedding dimensions are now mapped to provider settings.
Capability declarations now use confirmed model IDs, advertise file and mask support where applicable, and allow asynchronous lookups and middleware overrides to resolve to unknown.
Structured output defaults to the JSON tool fallback for Claude Sonnet 4.6 and Claude Haiku 4.5, while strict tool support is preserved.
Amazon Bedrock streaming responses now accept citation deltas.
Fixes and other changes 22
SigV4 signing is skipped for non-ASCII header values.
Structured output for Claude Haiku 5.5 on Amazon Bedrock uses the JSON tool fallback.
Partial configurations continue to preserve explicit values and derive missing fields; explicit effort or budget overrides and disabled thinking no longer emit warnings for portable mappings that are not sent.
Nova 2 calls using portable xhigh now send maxReasoningEffort: 'high' instead of the unsupported 'max', and return a compatibility warning.
Hostname parts are validated using isValidHostnamePart.
The User-Agent header now follows standards, improving compatibility with systems that inspect it.
The provider omits max output tokens for Nova 2 Lite high reasoning.
Unsupported OpenAI sampling settings are omitted from Converse requests.
Supported Mantle models are routed through the OpenAI endpoint.
Unsupported sampling parameters are omitted for newer Claude models.
Avoids forcing structured-output tools for incompatible Claude models.
API error messages are preserved when the error type is omitted.
Open tool schemas no longer use strict mode.
Anthropic features are enabled for application inference profile chat models.
Bedrock endpoints now support non-standard AWS regions and service-specific overrides.
Returns text from citation content responses.
Preserves provider-executed tool results when replaying messages.
Assistant messages are omitted when filtering unsigned reasoning leaves only cache points.
Amazon Bedrock now forwards the Anthropic option for disabling parallel tool use without sending conflicting tool choice fields.
Modeled event-stream exceptions are now surfaced by the Amazon Bedrock provider.
Handles empty string tool call IDs.
Supports reasoningContent.redactedContent from the Converse API and replays it on subsequent turns.
@ai-sdk/amazon-bedrock 49 changes
Breaking changes 2
Amazon Bedrock event stream frame decoding and processing failures are surfaced instead of silently completing the stream.
Amazon Bedrock event streams ending with an incomplete buffered frame are rejected instead of completing normally with partial output.
New APIs and options 1
Adds a structured output mode option.
Behavior and default changes 2
Native structured output is used for known supported Claude models; future Claude model IDs use JSON instructions for structured output by default.
Claude Sonnet 4.6 and Claude Haiku 4.5 default to the JSON tool fallback while preserving strict tool support.
Fixes and other changes 4
Strict-tool fields are omitted for future Claude model IDs that do not support them.
Generated text is now returned from citation content blocks.
Amazon Bedrock now handles channel-qualified JSON response tool names.
Raw usage metadata now preserves complete Amazon Bedrock Converse usage objects.
@ai-sdk/amazon-bedrock 39 changes
Breaking changes 1
For claude-sonnet-5-5, disabled thinking is replaced with between_tools thinking, and budget-based thinking is converted to adaptive thinking; disabled thinking emits a warning.
New APIs and options 2
Adds Claude Haiku 5.5 as a supported Amazon Bedrock model.
Adds the between_tools thinking type, the lowest thinking setting for claude-sonnet-5-5.
Behavior and default changes 3
For claude-sonnet-5-5, jsonTool falls back to native structured outputs with a warning.
For claude-sonnet-5-5, required and named tool choices fall back to auto with a warning because the model rejects forced tool use.
Claude Sonnet 4.6 and Claude Haiku 4.5 now default to the JSON tool fallback.
Fixes and other changes 3
For claude-sonnet-5-5, xhigh and max effort are lowered to high with a warning because the API rejects them with between_tools.
Resolves endpoints for non-standard AWS partitions and endpoint overrides.
Non-streaming citation responses now return text.