Blab AI
Amazon Bedrock Unveils Major Agent, Security, and Infrastructure Enhancements in August 2026
Amazon Web Services (AWS) rolled out a sweeping set of updates to its Bedrock platform in August 2026, extending the reach of its foundational AI services and the AgentCore framework. The enhancements bring new web‑search integration for OpenAI models, million‑token context windows, cross‑region inference, cost‑control tools, and upgraded agent runtime capabilities. In addition, AWS added security‑focused models Daybreak Blue and Red, support for regulated workloads, and a robotics extension through Strands Robots.
The most noticeable change is Bedrock’s Web Search tool for OpenAI models. When enabled, GPT‑5.6 Sol, Terra, and Luna can pull up‑to‑date information from the public web during a request and embed citations directly into the response. Bedrock handles the entire integration, so developers need only add a single tool to an API call to receive a grounded answer without managing external search services. The feature is available for all three GPT‑5.6 variants and is supported by the same Responses API that Bedrock uses for other foundation models.
Context handling has also been enlarged. GPT‑5.6 Sol, Terra, and Luna now support million‑token context windows and prompt caching. The larger windows allow a single request to reference an entire codebase or regulatory document, while caching cuts repeated token costs and reduces latency. Together, the expanded context and web search let a single API call analyze a full document, compare it with current public data, and return a cited answer.
Infrastructure improvements focus on scalability and cost visibility. Cross‑region inference is now available for GPT‑5.6 models in more than 25 AWS regions. Users can choose a global profile that routes requests across all regions for higher throughput, or a geographic profile that keeps inference within a specified region to satisfy data‑processing requirements. AWS also introduced IAM principal cost allocation, which attributes inference spend to a specific user or project, and Cost Anomaly Detection, which monitors foundation‑model spend and delivers root‑cause analysis when costs shift unexpectedly. OpenAI announced lower pricing for the GPT‑5.6 family on Bedrock.
AgentCore received several functional upgrades. Runtime instances can now run on dedicated EC2 compute—including GPU‑accelerated, memory‑optimized, and compute‑optimized instances—and sessions can last up to 14 days. Temporal policies let developers enforce action sequences, prerequisites, and data freshness checks on each agent action. Rate limiting controls requests, inference tokens, and concurrent connections by user or group. AgentCore payments enable agents to access and pay for APIs and paid content while enforcing spending limits and providing end‑to‑end observability. Web Search is also available within AgentCore, allowing agents to retrieve current information from approved domains.
Security and regulated‑workload support were expanded with Daybreak Blue and Red. Daybreak Blue supports defensive tasks such as vulnerability discovery and incident response, while Daybreak Red enables authorized research and exploit reproduction. Both models are available on Bedrock for eligible customers. Claude Opus 5 and GPT‑5.6 Terra and Luna are now offered in AWS GovCloud (US) regions with zero data retention enabled by default. Amazon Nova Multimodal Embeddings is also available in GovCloud, permitting retrieval across text, images, video, and audio.
The robotics extension, Strands Robots, connects Strands Agents, LeRobot, and Hugging Face Storage Buckets. The workflow supports recording demonstrations, streaming datasets in the LeRobot format, training policies, and deploying them to simulated or physical hardware without data conversion. Strands Robots also supports mesh‑based discovery across simulated and physical devices, using Zenoh for local networks and AWS IoT Core for geographically distributed fleets. The extension participates in a limited research preview of the Model Hardware Standard.
Overall, the August 2026 updates broaden Bedrock’s reach into enterprise, regulated, and physical domains while giving developers tighter control over context, cost, and agent behavior. The changes position Bedrock as a more complete platform for building, scaling, and governing AI systems across a wide range of workloads.
The most noticeable change is Bedrock’s Web Search tool for OpenAI models. When enabled, GPT‑5.6 Sol, Terra, and Luna can pull up‑to‑date information from the public web during a request and embed citations directly into the response. Bedrock handles the entire integration, so developers need only add a single tool to an API call to receive a grounded answer without managing external search services. The feature is available for all three GPT‑5.6 variants and is supported by the same Responses API that Bedrock uses for other foundation models.
Context handling has also been enlarged. GPT‑5.6 Sol, Terra, and Luna now support million‑token context windows and prompt caching. The larger windows allow a single request to reference an entire codebase or regulatory document, while caching cuts repeated token costs and reduces latency. Together, the expanded context and web search let a single API call analyze a full document, compare it with current public data, and return a cited answer.
Infrastructure improvements focus on scalability and cost visibility. Cross‑region inference is now available for GPT‑5.6 models in more than 25 AWS regions. Users can choose a global profile that routes requests across all regions for higher throughput, or a geographic profile that keeps inference within a specified region to satisfy data‑processing requirements. AWS also introduced IAM principal cost allocation, which attributes inference spend to a specific user or project, and Cost Anomaly Detection, which monitors foundation‑model spend and delivers root‑cause analysis when costs shift unexpectedly. OpenAI announced lower pricing for the GPT‑5.6 family on Bedrock.
AgentCore received several functional upgrades. Runtime instances can now run on dedicated EC2 compute—including GPU‑accelerated, memory‑optimized, and compute‑optimized instances—and sessions can last up to 14 days. Temporal policies let developers enforce action sequences, prerequisites, and data freshness checks on each agent action. Rate limiting controls requests, inference tokens, and concurrent connections by user or group. AgentCore payments enable agents to access and pay for APIs and paid content while enforcing spending limits and providing end‑to‑end observability. Web Search is also available within AgentCore, allowing agents to retrieve current information from approved domains.
Security and regulated‑workload support were expanded with Daybreak Blue and Red. Daybreak Blue supports defensive tasks such as vulnerability discovery and incident response, while Daybreak Red enables authorized research and exploit reproduction. Both models are available on Bedrock for eligible customers. Claude Opus 5 and GPT‑5.6 Terra and Luna are now offered in AWS GovCloud (US) regions with zero data retention enabled by default. Amazon Nova Multimodal Embeddings is also available in GovCloud, permitting retrieval across text, images, video, and audio.
The robotics extension, Strands Robots, connects Strands Agents, LeRobot, and Hugging Face Storage Buckets. The workflow supports recording demonstrations, streaming datasets in the LeRobot format, training policies, and deploying them to simulated or physical hardware without data conversion. Strands Robots also supports mesh‑based discovery across simulated and physical devices, using Zenoh for local networks and AWS IoT Core for geographically distributed fleets. The extension participates in a limited research preview of the Model Hardware Standard.
Overall, the August 2026 updates broaden Bedrock’s reach into enterprise, regulated, and physical domains while giving developers tighter control over context, cost, and agent behavior. The changes position Bedrock as a more complete platform for building, scaling, and governing AI systems across a wide range of workloads.
