AWS announced a public preview of OpenAI-powered managed Amazon Bedrock Agents, enabling developers to build agents optimized for OpenAI models and run them entirely within AWS while leveraging the identities, permissions, and governance controls already used in their accounts.
The service is based on a customized version of OpenAI’s Agents API, designed to integrate with the AWS environment and its resources. Customers can choose to run the agents on compute infrastructure they manage themselves, such as a development machine, container, or existing compute environment, or use Amazon Bedrock AgentCore Runtime to provide managed runtime sessions and configurable storage within their AWS account.
New models on Amazon Bedrock
AWS added a group of models to Bedrock. The list includes OpenAI GPT-6.1 Sol, which OpenAI says delivers improved performance in agentic coding, computer use, and professional tasks, with performance approaching GPT-6 Astra on difficult tests and at roughly one-fifth the cost, according to the company.
GPT-6 Astra is also now available in UltraFast mode, a premium speed tier that OpenAI says can provide up to six times faster inference, at speeds of up to 300 tokens per second. The additions also include Anthropic Claude Sonnet 5.5, which targets improved coding and execution of defined tasks within larger workflows, and SpaceXAI Grok 4.7, with improvements in handling mixed documents, repository-level coding, and browser-use agents.
Other launches relevant to infrastructure and data teams
AWS also introduced a preview version of AWS Well-Architected Agent, an agent that analyzes an AWS environment and suggests recommendations related to cost, security, performance, and resilience while taking the organization’s goals and infrastructure context into account.
Amazon Aurora PostgreSQL can now query Apache Iceberg and Parquet data directly alongside operational data, using existing PostgreSQL applications and tools without the need for ETL pipelines or additional data copies. Amazon S3 Tables also added support for all Apache Iceberg V3 data types, including geometry, geography, nanosecond-precision timestamps, and column default values.
What changes in practice?
These updates combine expanded model options with keeping agent execution within existing AWS controls, which may reduce operational complexity for teams that want to use OpenAI models with their internal resources and permissions. However, the announcement remains in the public-preview stage for Managed Agents, and the performance and cost claims for the models are attributed to their developers rather than being independent results in the material.
Changes to the services life cycle
AWS pointed to life-cycle updates published on September 29, 2026. Amazon Chime SDK SIP Media Application and Amazon WorkSpaces Secure Browser will move into maintenance mode, preventing new customers from accessing them beginning October 29, 2026.
Amazon Managed Blockchain, Amazon DevOps Guru, and AWS Backint Agent for SAP ASE have also entered the sunset phase, with support ending on September 29, 2027, September 30, 2027, and September 29, 2027, respectively. Support for the standalone AWS Infrastructure Composer console ends on December 7, 2026, while Amazon Mechanical Turk reached end of support on September 29, 2026.
Other news included the introduction of Kiro workflows for executing complex tasks across multiple agents with less supervision, the launch of Strands Decider 2B as an open-source decision model aimed at rapid experimentation and local development, and three new certification tracks for AWS partners under the Partner-Led FDE initiative.