Amazon Bedrock changelog

Terra and Luna models on bedrock-runtime with Responses, Converse, Chat Completions APIs and Cross-Region Inference in G

GPT-5.6 Cyber for vulnerability research offered to eligible customers in GovCloud (US-East)Region

Standard and Express workflows usable as AI agent tools through AgentCore Gateway

Rebuilt with faster startup, consistent commands, and support for more AgentCore resourcesGA

Added in London, Frankfurt, Ohio and Canada (Central), for nine Regions totalRegion

Canada cross-Region inference routes only between Canada Central and Canada West

1 million token context, text and image input, token prices at least 50% lower than GPT-6.1 SolGA

Domain filtering of up to 100 allowed or blocked domains and user location controls added

Supports InvokeModel APIs and model fallback routing with a circuit breaker on its inference endpoint

Schema-compliant JSON responses for all Claude models on bedrock-runtime and bedrock-mantle endpoints

Standard tier for contextual grounding checks supports conversational and reasoning responses, in 11 Regions

Optional prompt shortening returns shorter prompts that score at least as high

Available in AWS GovCloud (US) for coding, agentic tasks and knowledge workGA

Expands to US East (Ohio), US West (Oregon), Asia Pacific (Sydney) and GovCloud (US-West)Region

reasoning.summary parameter supported for OpenAI models through the Responses API

Video-to-text model returns structured, time-coded metadata from video, via US and Global inference profilesGA

Bedrock costs can be grouped and filtered by model, provider, inference type and feature

Fastest Claude 5.5 model, offered on Amazon Bedrock and Claude Platform on AWSGA

Fastest Claude 5.5 model offered through Amazon Bedrock in AWS GovCloud (US)GA

Z.ai open weight model with 1M-token context and prompt cachingGA

Supports private CA-signed TLS certificates on MCP, OpenAPI and HTTP proxy targets in VPC

Premium speed tier with up to 6x faster inference, up to 300 tokens per secondGA

In-region inference added in UK (London) Region, keeping data processing in the UKRegion

OpenAI-powered agents run inside AWS with IAM roles, durable sessions and MCP toolsPreview

Upgrade to GPT-6 Sol for agentic coding and computer use, with explicit prompt cachingGA

Claude Opus 5, Sonnet 5 and Haiku 4.5 offered in India, South Korea and SingaporeRegion

Coding and agentic model with US Geo and Global cross-Region inferenceGA

Offered through Amazon Bedrock and Claude Platform on AWS, with stronger coding and knowledge workGA

Offered in AWS GovCloud (US) through Amazon BedrockGA

Native Salesforce and Zendesk data source connectors sync articles and community posts

Both models on Bedrock with up to 1M tokens of contextGA

Available on Amazon Bedrock and Claude Platform on AWSGA

Offered on AWS GovCloud (US) through Amazon BedrockRegion

Moonshot AI's 2.8 trillion parameter open model with vision, 1M-token context and prompt cachingGA

Next-generation Runtime reclaims unused memory and gives P75 cold starts of 1.9 to 2.0 secondsGA

OpenAI's latest model on Bedrock, with a context window of up to 1 million input tokensGA

Available in AWS GovCloud (US-West) for GPT-5.4 and GPT-5.6 Terra and LunaRegion

Managed consent portal removes need for custom OAuth callback infrastructure for AgentCore Gateway

Anthropic's frontier model available through Amazon Bedrock and Claude Platform on AWSGA

Anthropic's frontier model on Amazon Bedrock in AWS GovCloud (US)GA

Private catalog for agents, tools and MCP servers adds IaC support, RAM sharing and auto-detectionGA

Available on Amazon Bedrock in AWS GovCloud (US) with 500k context windowGA

Available in US West (N. California) and Asia Pacific (Hyderabad)Region

Available on Bedrock in GovCloud (US-West) and (US-East) with 1 million token contextGA

Now $4 per million input and $20 per million output tokens, through at least November 21, 2026Price

Adds per-request domain and published-date filtering, and expands to Europe (Ireland) and Asia Pacific (Tokyo)

AI agents can pay for APIs, MCPs and content, with Coinbase and Stripe Privy walletsGA

Available in AWS GovCloud (US) Regions through Amazon Bedrock with zero data retentionRegion

Cyber defense models GPT-5.6 Cyber and GPT-5.6 Sol for eligible customers in US East (Ohio)GA

Available in AWS GovCloud (US-West)Region

IAM principal cost allocation now covers the bedrock-mantle endpoint

Memory, policy and managed harness added in AWS GovCloud (US-West)Region

Run agents on EC2 instances through capacity providers, with sessions up to 14 daysGA

Adds temporal policies for stateful agent authorization and per-user rate limiting for gateway traffic

Built-in server-side web search for OpenAI GPT models, with no data leaving AWSGA

1 million token context windows supported on Bedrock, with prompt caching discount

Open source tool drafts business ontologies for AI agents, stored in a Neptune knowledge graphGA

On-demand prices drop 80% for Luna and 20% for Terra from July 30, 2026Price

31B, 26B-A4B and E2B variants available in AWS GovCloud (US-West)Region

Available in AWS GovCloud (US-West), adding xAI as a model provider thereRegion

Anthropic's most advanced Opus model, offered via Amazon Bedrock and Claude Platform on AWSGA

Agent traces, prompts and logs delivered to a single per-agent CloudWatch log group

Available on Amazon Bedrock in AWS GovCloud (US) via bedrock-runtime and bedrock-mantle endpointsRegion

Standardized Bedrock product metadata including model provider, model name and inference type

Default runtime quotas raised to 5,000 concurrent sessions in US East (N. Virginia) and US West (Oregon), 2,500 elsewher

Available in Bangkok, Malaysia, Milan and Spain RegionsRegion

Anthropic's latest Sonnet model offered via Amazon Bedrock and Claude Platform on AWSGA

Offered in AWS GovCloud (US) through Amazon BedrockGA

FedRAMP High and DoD IL-4/5 approved in AWS GovCloud (US) Regions

Cross-account access to memory resources and delivery destinations via resource-based policies

Two automated policy refinement workflows: iterative improvement and ambiguity reduction

Managed agent harness defines agents in configuration, exports to Strands-based codeGA

Insights, recommendations, batch evaluation and A/B testing to improve agents from production tracesGA

Bedrock Guardrails evaluate agent action outputs and gateway target inputs in policyGA

Managed RAG service with six data source connectors and agentic retrievalGA

New resourceless InvokeGuardrailChecks API applies individual safeguards in agent loops, detect-only

Automated Reasoning checks added in Asia Pacific (Sydney) RegionRegion

Managed web search tool for agents, exposed via AgentCore gateway over MCP, in US East (N. Virginia)GA

Strictly consistent metadata passes application-supplied values to long-term records without LLM inference

Mythos-class model available on Amazon Bedrock and Claude Platform on AWSGA

Interactive shells give terminal access into agent sessions, up to 10 concurrent shells

Redesigned console for the bedrock-mantle endpoint with model comparison, projects and prefilled code samples

Available in AWS GovCloud (US-West)Region

Credential Providers can reference existing AWS Secrets Manager secret ARNs, in 14 RegionsGA

CloudWatch metrics for inference counts, token totals and client errors

Models and Codex run on Bedrock at OpenAI first-party pricingGA

Anthropic's most capable generally available model, offered via Amazon Bedrock and Claude Platform on AWSGA

Inference quotas for the bedrock-mantle endpoint now visible in AWS Service Quotas

Request-level usage attribution added to InvokeModel and InvokeModelWithResponseStream APIs

Optimizes prompts for any Bedrock model and compares results across up to 5 models

AI agents pay for APIs, MCP servers and content via Coinbase or Stripe wallets with spending limitsPreview

Mounts Amazon S3 Files and Amazon EFS access points as bring-your-own file systems

Available in AWS GovCloud (US-West)Region

Available in South America (São Paulo) Region with runtime, identity, gateway, policy, observability, code interpreter aRegion

Connects AI agents to SAP ERP systems on AgentCore Runtime using MCP and ODataGA

On-Behalf-Of token exchange lets agents access resources for users without repeated consent flowsGA

Adds recommendations, batch evaluations and A/B tests to validate agent performance changesPreview

Node.js supported for direct code deployment from a .zip file, alongside Python

Limited preview of OpenAI models, Codex and Managed Agents powered by OpenAIPreview

Adds managed harness (preview), AgentCore CLI and skills for coding assistantsPreview

Anthropic's most capable Opus model, served by Bedrock's next-generation inference engineGA

Cost allocation by IAM user and role in CUR 2.0 and Cost Explorer

Private catalog for discovering and governing agents, tools, skills and MCP servers via AgentCorePreview

OS-level mouse, keyboard and full desktop screenshot actions beyond Chrome DevTools Protocol

Gated research preview in US East (N. Virginia), limited to allow-listed organizationsPreview

Custom vocabulary via Data Automation Library improves transcription of domain terms in 11 languages, at no extra charge

Cross-account safeguards enforce a management account guardrail across all organization accounts and OUsGA

Automated quality assessment for AI agents with 13 built-in evaluators, in nine RegionsGA

Chrome Enterprise policies for Browser and custom root CA for Browser and Code Interpreter

Managed session storage keeps agent filesystem state across stop and resume, up to 1 GBPreview

WebRTC added for real-time bidirectional streaming, in fourteen RegionsGA

Two agent-focused models added in select AWS RegionsGA

Open hybrid MoE model for multi-agent applications added to Bedrock in select RegionsGA

New InvokeAgentRuntimeCommand API runs shell commands in a session, streaming output over HTTP/2

Supports the AG-UI protocol for deploying real-time agent servers in fourteen Regions

Streaming notifications to Kinesis when long-term memory records are created or modified

Stateful MCP servers support elicitation, sampling and progress notifications in 14 Regions

Centralized agent-tool access controls written in natural language converted to Cedar, in 13 RegionsGA

Supports Converse API as a model invocation type for batch jobs

OpenAI-compatible projects isolate applications, with per-project IAM access control and tags in Mantle

AgentCore Gateway tools run server-side in a single Responses API callGA

Open-weight coding model available through the Bedrock managed API in select RegionsGA

Adds open-weight models Qwen3 32B and GPT-OSS 20B, plus OpenAI-compatible fine-tuning APIs

Anthropic model for coding, agents and computer use, upgrade from Sonnet 4.5GA

Open-weight models supported in Asia Pacific (Sydney) via bedrock-runtime and bedrock-mantle endpointsRegion

PrivateLink now covers the OpenAI-compatible bedrock-mantle endpoint in 14 listed Regions

Supports customer-provided HTTP and HTTPS proxy configuration for browser sessions

Amazon Bedrock adds support for six fully-managed open weights models

Amazon Bedrock AgentCore Browser now supports browser profiles

Claude Opus 4.6 now available in Amazon Bedrock

Structured outputs now available in Amazon Bedrock

Amazon Bedrock now supports server-side custom tools using the Responses API

Amazon Bedrock now supports 1-hour duration for prompt caching

Amazon Bedrock AgentCore Browser now supports custom browser extensions

Amazon Bedrock Reserved Tier available now for Claude Sonnet 4.5 in AWS GovCloud (US-West)

Amazon Bedrock Reserved Tier available now for Claude Opus 4.5 and Haiku 4.5

AWS Data Exports adds granular operation visibility for Amazon Bedrock model usage

Amazon Bedrock AgentCore Policy now supports Bedrock Guardrails in US West (Oregon) and US East (Ohio)

Amazon Bedrock introduces API keys to streamline development in GovCloud regions