AWS September recap puts model choice, agent efficiency and enterprise data at the center

Amazon Web Services used its September 2026 product recap to frame enterprise AI development around more than model performance. The company highlighted expanded model access in Amazon Bedrock, lower-overhead execution through Amazon Bedrock AgentCore, new open-source tools from Strands, and automated synchronization for enterprise knowledge bases.

The update matters because AI builders are increasingly deploying agents that must retrieve company information, use business systems, and operate over extended tasks. AWS’s own account of the month’s releases suggests that decisions about latency, token consumption, permissions, data freshness, and operational controls are becoming as important as choosing a high-performing model.

Amazon Bedrock broadens the model menu

AWS said OpenAI’s Astra, Sol, and Luna model families became generally available through Amazon Bedrock during September. The company describes GPT-6 Astra as a high-end option for complex decisions, document analysis, and software development, with an input context window of up to 1 million tokens. GPT-6 Astra Ultrafast is positioned as a faster tier for workloads where response speed is a priority.

The Bedrock lineup also includes GPT-6.1 Sol and GPT-6 Sol for recurring coding, computer-use, and professional workloads, while GPT-6.1 Luna is aimed at high-volume activities such as extraction, classification, summarization, and routing. AWS presents the grouping as a way for teams to balance intelligence, latency, throughput, and cost rather than using one model for every task.

AWS also listed new Claude options, including Claude Fable 5.1, Claude Opus 5.5, and Claude Sonnet 5.5. The company said Opus 5.5 is intended for agentic coding, knowledge work, and long-running tasks, while Sonnet 5.5 offers a more efficient option for focused coding and knowledge work. Claims that Sonnet 5.5 delivers 30 percent lower cost per task and 30 percent faster speed than Claude Sonnet 5 come from the AWS product recap and were not independently verified in the source material.

Moonshot AI’s Kimi K3 and xAI’s Grok 4.6 and Grok 4.7 were also added to the catalog, according to AWS. The company cited Moonshot’s description of Kimi K3 as a 2.8 trillion-parameter open model and reported that the Grok models support a 500,000-token context window, configurable reasoning effort, and self-verification.

Agent infrastructure moves toward usage-based execution

For teams building production agents, AWS emphasized changes below the model layer. Amazon Bedrock Managed Agents powered by OpenAI entered public preview, allowing developers to use OpenAI models while keeping data within AWS, reusing AWS Identity and Access Management permissions, and recording activity through AWS CloudTrail. AWS also cited durable sessions and human-approval workflows as controls designed to support higher-risk deployments.

The latest AgentCore runtime is designed to reduce infrastructure overhead for serverless agents. AWS said it improves memory management and cold-start latency, scales sessions to zero when idle, and uses hardware-isolated environments. The pricing model is described as pay-as-you-go, with customers charged for actual usage rather than pre-provisioned peak memory.

Strands, AWS’s open-source agent toolkit, gained a new Strands harness. AWS claims the harness matches popular alternatives on accuracy while using 28 percent fewer tokens. The harness supports Python and TypeScript, with context management, prompt caching, and memory included, and is intended to run across deployment environments.

AWS also introduced Strands Decider 2B, a 2-billion-parameter open-source model that selects from predefined options instead of generating text. The company says it can respond locally in about 115 milliseconds and is intended for tool selection, routing, and guardrails. The code, data, and weights are available through GitHub and Hugging Face, according to AWS.

Enterprise knowledge bases get more direct connectors

The September releases also target one of the less visible but persistent problems in enterprise AI: keeping retrieved information current without maintaining large custom ingestion systems.

Amazon Bedrock Managed Knowledge Base now supports daily, weekly, and monthly automatic refresh schedules for all native data-source connectors, AWS said. User-managed configurations for SharePoint, OneDrive, and Confluence are intended to reduce dependence on administrator-managed service accounts when an organization already has the necessary access.

New native connectors for ServiceNow, Confluence Data Center, Salesforce, and Zendesk can handle crawling, metadata extraction, and incremental synchronization. AWS says these capabilities reduce custom ingestion code for internal documentation, support content, and operational knowledge. In practice, the value will depend on connector coverage, permissions, sync reliability, and how accurately changes are reflected in agent responses—areas the recap does not quantify.

What the evidence means for builders and buyers

The strongest claims in this update are vendor-reported. The source cluster contains an AWS wire listing and an AWS Machine Learning Blog post, but no independent testing, customer interviews, pricing comparisons, or third-party adoption data. Assertions about token savings, inference speed, model capability, and cost efficiency should therefore be treated as product positioning until validated against a team’s own workload.

Even with that limitation, the product direction is clear. Builders can use Bedrock’s wider catalog to route tasks by complexity and cost, while AgentCore and Strands address execution, memory, approvals, and deployment. For enterprise buyers, native connectors and scheduled synchronization may reduce maintenance work, but they do not remove the need to evaluate access controls, data lineage, retention, and failure handling.

The practical test will be whether these components improve total system economics rather than isolated benchmark results. A faster runtime may have limited value if retrieval is stale. A cheaper model may create review costs if routing or extraction quality falls. Likewise, a managed connector can simplify deployment while still requiring careful permission design around sensitive systems such as Salesforce, ServiceNow, and Zendesk.

What to watch next

Builders should watch for general availability dates and production pricing for the OpenAI-powered Managed Agents preview, along with documentation on regional availability and supported model features. Independent measurements of AgentCore cold starts, Strands token use, and Strands Decider 2B accuracy would help establish how the reported improvements translate to real workloads.

Enterprise teams should also test the new knowledge-base connectors against permission changes, deleted records, incremental updates, and conflicting source documents. The next meaningful signal will be evidence that these connectors keep agent context current without introducing access or governance gaps.

Creati.ai perspective

AWS’s September recap is less a single model announcement than a coordinated push to make agent systems easier to assemble and operate. Its emphasis on model choice, managed execution, local decision-making, and enterprise data synchronization reflects where production friction now sits: in the surrounding system, not only in the language model.

For AI teams, the opportunity is broader choice, but the trade-off is a more complex evaluation matrix. AWS’s claims provide a useful roadmap of what to test, not a substitute for testing. The buyers best positioned to benefit will measure end-to-end cost, reliability, freshness, and governance across real workflows rather than selecting models on headline capability alone.