ICYMI: What landed for AI builders in September 2026

AWS published a September 2026 roundup of updates across Amazon Bedrock, Amazon Bedrock AgentCore and Strands, describing the changes as expanding model choice, improving agent runtime efficiency and tightening connections between AI applications and current business data.

On the agent tooling side, Amazon Bedrock Managed Agents powered by OpenAI entered public preview. AWS says builders can use OpenAI models while keeping data within AWS, reuse existing AWS Identity and Access Management permissions, and retain auditability through AWS CloudTrail. The service also offers durable sessions and built-in human approval workflows, which AWS frames as reducing development overhead and risk.

The AgentCore runtime received more efficient memory management and lower cold start latency for serverless agents. AWS's pitch is that customers pay for actual usage rather than peak memory; sessions scale to zero when idle, run in hardware-isolated environments, and are billed pay-as-you-go with no pre-provisioned capacity. This suggests AWS is competing on the operational economics of agent hosting, not just model access.

Two Strands releases target token cost and local decision-making. Strands harness is a new open source agent harness that AWS says matches popular harnesses on accuracy while using 28 percent fewer tokens; a production-ready agent can be started in a single line of Python or TypeScript, with built-in context management, prompt caching and memory, and deployed anywhere. Strands Decider 2B is a 2-billion-parameter open source decision model that selects among predefined options rather than generating text, returning answers in roughly 115 milliseconds locally. It is aimed at agentic tasks such as tool selection, routing and guardrails, with code, data and weights on GitHub and Hugging Face. The practical read is that routing and guardrail decisions can run on-device instead of round-tripping to a hosted model.

Model availability saw the largest set of changes. OpenAI's Astra, Sol and Luna families are now generally available on Bedrock. GPT-6 Astra is positioned as the flagship for demanding work including complex decisions, document analysis and software development, with up to 1 million input tokens; AWS describes it as operating across software and files and producing output aligned with organizational voice, templates and standards. GPT-6 Astra Ultrafast is a premium speed tier delivering up to 6x faster inference in the API at up to 300 tokens per second. Sol targets coding, computer use and frequently run professional workloads, with both GPT-6.1 Sol and GPT-6 Sol available. GPT 6.1 Luna is aimed at high-volume extraction, summarization, classification and routing.

Anthropic's Claude line expands with Fable 5.1 for coding, scientific research and enterprise workflows, plus Opus 5.5 for agentic coding, knowledge work and long-running tasks. Opus 5.5 uses adaptive thinking to gauge how much reasoning a task needs and an effort parameter that sets an upper bound on reasoning depth. Sonnet 5.5 is described as 30 percent lower cost per task and 30 percent faster than Claude Sonnet 5 for focused coding and knowledge work. The explicit effort parameter is a notable design choice: it gives teams a direct dial on the cost-versus-thoroughness trade-off, which matters when agents run unattended.

Moonshot AI's Kimi K3 is now available on Bedrock. According to Moonshot AI, it is the company's most capable model and the first open model to reach 2.8 trillion parameters, with a 1-million-token context window, native vision and built-in prompt caching; AWS says it helps build apps 2.5x faster at lower cost than previous versions. xAI's Grok 4.6 and Grok 4.7 offer a 500K token context window and support configurable reasoning effort plus self-verification for long-running tasks. Taken together, the portfolio now spans extremes of context length, throughput and price, which shifts the hard part of building from picking the best model to matching each workload to the right one.

Knowledge grounding also changed. Amazon Bedrock Managed Knowledge Base added automatic sync scheduling, letting teams configure daily, weekly or monthly refresh options for all native data-source connectors so agents retrieve current information. 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, cutting the custom ingestion pipelines teams previously had to build and maintain for support content, internal documentation and operational knowledge. In practice, native connectors plus scheduled sync remove a common source of staleness and maintenance burden, though teams should still expect to review permission mapping and sync frequency against their own data-governance rules.

AWS points builders to Amazon Bedrock, the AgentCore CLI and the Strands Harness SDK as starting points.

Why it matters: Teams standardizing on Bedrock now have a much wider model menu from OpenAI, Anthropic, Moonshot AI and xAI inside one platform, which reduces the need to juggle multiple vendors for different workloads. The runtime and harness changes lower the cost floor for running agents, and the knowledge base connectors reduce the glue code needed to keep them current. The main follow-up work for developers is evaluation: with this many models and controls like reasoning effort, choosing per workload becomes an ongoing tuning exercise rather than a one-time decision.