AI news ·
Amazon Bedrock expands model access, adds faster agent runtimes, and syncs enterprise knowledge automatically
AWS rolled out AI agent updates in September 2026, including a new open-source harness using 28% fewer tokens and a 2B-parameter decision model running in 115ms.
Amazon Web Services rolled out a series of updates across its AI development stack in September 2026, expanding model choice, reducing agent latency, and adding native connectors for enterprise data sources like ServiceNow and Salesforce. The changes target teams moving AI agents into production workflows where security, governance, and cost-per-task now carry as much weight as raw model performance.
"As AI models become more capable and model choice expands, the industry conversations are shifting. Model performance is no longer the only question," AWS said. "Customers now weigh cost against benefit for their specific use case, and those decisions matter when delegating work to agents in production."
Faster agents with lower overhead
Amazon Bedrock Managed Agents now supports OpenAI models in public preview. Teams can build agents with OpenAI models while keeping data inside AWS, reusing existing IAM permissions, and maintaining audit trails through CloudTrail. The service includes durable sessions and built-in human approval workflows.
AgentCore's updated runtime reduces cold start latency for serverless agents through more efficient memory management. Sessions scale to zero when idle and run in hardware-isolated environments with pay-as-you-go pricing - no pre-provisioning required.
Strands harness, a new open source agent harness, matches popular alternatives on accuracy while using 28 percent fewer tokens. Developers can spin up a production-ready agent in a single line of Python or TypeScript, with built-in context management, prompt caching, and memory, then deploy it anywhere. A companion model, Strands Decider 2B, is a 2-billion-parameter decision model that picks between predefined options in roughly 115 milliseconds locally, targeting tool selection, routing, and guardrail tasks.
Expanded model catalog for different workloads
OpenAI's Astra, Sol, and Luna models reached general availability on Amazon Bedrock. GPT-6 Astra handles complex decisions, document analysis, and software development with up to 1 million input tokens. GPT-6 Astra Ultrafast adds a premium speed tier delivering up to 300 tokens per second - roughly six times faster inference. Sol targets coding and professional workloads that run frequently, while GPT 6.1 Luna is designed for high-volume extraction, summarization, classification, and routing.
Anthropic's Claude lineup also expanded. Claude Opus 5.5 uses adaptive thinking to gauge how much reasoning a task needs, with an effort parameter to set an upper bound on reasoning depth. Claude Sonnet 5.5 delivers 30 percent lower cost per task at 30 percent faster speed than its predecessor. Claude Fable 5.1 rounds out the options for coding and scientific research.
Moonshot AI's Kimi K3 - a 2.8-trillion-parameter open model with a 1-million-token context window, native vision, and built-in prompt caching - is now available alongside xAI's Grok 4.6 and Grok 4.7, which offer a 500K token context window and configurable reasoning effort with self-verification.
Enterprise data connectors and knowledge sync
Amazon Bedrock Managed Knowledge Base added automatic sync scheduling with daily, weekly, and monthly refresh options for all native data-source connectors. This keeps agent knowledge current without manual intervention. The user-managed setup for SharePoint, OneDrive, and Confluence reduces reliance on admin-managed service accounts.
New native connectors for ServiceNow, Confluence Data Center, Salesforce, and Zendesk handle data crawling, metadata extraction, and incremental sync automatically. These reduce the amount of custom ingestion code teams need to maintain for support content, internal documentation, and operational knowledge. For developers building agent workflows that depend on current business information, AI Engineering Courses cover the patterns that make these integrations work in production.
Why this matters for operations and IT teams
The September updates address a practical shift: agent performance is now measured by the system around the model, not just the model itself. For operations and IT teams managing production agents, the combination of faster cold starts, native enterprise connectors, and granular model choices - from a 2B-parameter local decider to trillion-parameter reasoning models - means you can match infrastructure cost to task complexity. Pay-as-you-go pricing with scale-to-zero sessions removes the need to over-provision for intermittent workloads, while automatic knowledge sync and CloudTrail auditability reduce the operational burden of keeping agents both current and compliant.