AWS AI Updates: September 30, 2026

1. Bedrock Managed Agents, Built With OpenAI, Enters Preview

AWS. AWS and OpenAI jointly built Amazon Bedrock Managed Agents (BMA), a customized, AWS-native version of OpenAI’s Agents API for building agents optimized for OpenAI models that run entirely inside AWS. BMA handles state, tool selection, code execution, and multistep coordination, with durable sessions that keep messages, tool calls, and intermediate results so a user can return later and continue. Agents can use reusable skills and MCP servers, each runs under its own IAM role, consequential actions can require human approval, and supported API activity is logged to CloudTrail. The preview has no extra charge beyond underlying resources and runs in US East (N. Virginia), US East (Ohio), and US West (Oregon). Source

2. GPT-6.1 Sol Is Generally Available on Bedrock

AWS. OpenAI’s GPT-6.1 Sol, an upgrade to GPT-6 Sol aimed at agentic coding, computer use, and professional work, is now generally available on Amazon Bedrock. AWS cites OpenAI results showing it matches GPT-6 Astra on DeepSWE v1.1 at roughly one-fifth the cost per task and beats GPT-6 Sol by 6.4 percentage points at a lower reasoning effort. It supports explicit prompt caching on Bedrock, PrivateLink VPC endpoints, and use from Codex (desktop app, CLI, and IDEs); inference data is not used for training, abuse detection data is kept up to 30 days, and zero data retention is available on request. Source

3. AWS Transform’s Assessment Agent Now Sizes Kafka-to-MSK Migrations

AWS. AWS Transform’s agentic migration assessments now cover moving on-premises Apache Kafka clusters to Amazon MSK. Users upload a cluster inventory file or describe clusters in chat (or through the AWS Transform MCP server), and the agent checks each cluster’s topology, version, configuration, authentication, and quotas for MSK compatibility, right-sizes MSK Express brokers, and projects costs across broker hours, storage, data-in, and cross-AZ transfer. What-if scenarios can compare Regions, retention settings, or alternative configurations to build a TCO business case. Source

4. Condé Nast Indexed 140,000 Videos With TwelveLabs Marengo on Bedrock

AWS. Condé Nast built semantic video discovery over more than 140,000 videos using TwelveLabs Marengo on Amazon Bedrock, which produces embeddings that jointly encode visual, audio, and transcript signals per video segment, indexed in Amazon OpenSearch Service for k-NN search with metadata filters. Ingestion and query run as separate planes, with asynchronous batch embedding through Step Functions for the backfill and a multi-AZ deployment on ECS Fargate. Condé Nast reports discovery tasks dropping from 250 minutes to about 2 minutes and estimates about $800,000 in annual savings; segment length was tuned through iterative experiments. Source

5. A Contract Extraction Pipeline Uses Claude Haiku to Check Claude Sonnet and Textract to Break Ties

AWS. A reference architecture on Amazon Bedrock AgentCore Runtime uses Strands agents in a dual-model pattern: Claude Sonnet 4.6 reads contract PDFs natively and emits JSON with confidence scores, Claude Haiku 4.5 independently verifies the fields, and Amazon Textract settles disagreements such as signature detection. Policy in AgentCore applies Cedar-based access controls, results land in Aurora PostgreSQL, and Amazon Quick provides dashboards plus a chat agent that answers both portfolio aggregates and per-document questions. At 1,000 contracts a month AWS estimates about $349 total, of which Bedrock inference is $12 and Quick is $290, or $0.014 of AI cost per contract. Source