AWS Updates Bedrock, AgentCore, and Strands With New Models and Runtimes
AWS has updated Amazon Bedrock, AgentCore, and open-source Strands, adding models from OpenAI, Anthropic, Moonshot AI, and xAI alongside faster agent runtimes.
Amazon Web Services has rolled out a broad set of infrastructure and model updates across Amazon Bedrock, Amazon Bedrock AgentCore, and the open-source Strands project. The releases reflect an evolving industry focus: while raw intelligence benchmarks remain important, development teams are increasingly prioritizing runtime latency, token consumption, operational governance, and direct integration with enterprise data sources when deploying production agents.
Expanded frontier models across Bedrock
Bedrock's managed catalog now includes general availability for several key model families, alongside public previews for new agent runtimes. Headlining the OpenAI lineup on Bedrock are GPT-6 Astra, which supports up to a 1-million-token input window for extensive document analysis and code reasoning, and GPT-6 Astra Ultrafast, a speed-focused variant delivering up to 300 tokens per second and up to 6x faster inference. For routine workloads, AWS added GPT-6 Sol and GPT-6.1 Sol for computer use and recurring coding, while GPT 6.1 Luna handles high-volume tasks such as data extraction, classification, and text routing.
OpenAI integration extends further with Amazon Bedrock Managed Agents, now available in public preview. This service enables teams to run OpenAI-backed agents inside AWS infrastructure, leveraging existing AWS Identity and Access Management (IAM) policies, AWS CloudTrail audit logs, durable sessions, and built-in human-in-the-loop approvals.
Anthropic's footprint on Bedrock expanded with three new Claude releases. Claude Opus 5.5 introduces adaptive thinking to balance reasoning depth on long-running tasks via an effort parameter. Claude Sonnet 5.5 delivers focused coding and knowledge automation at a 30 percent speed improvement and 30 percent lower cost compared to Claude Sonnet 5, while Claude Fable 5.1 addresses scientific and software development pipelines.
AWS also added Moonshot AI's Kimi K3, a 2.8-trillion-parameter open model featuring a 1-million-token context window, native vision processing, and prompt caching designed to run 2.5x faster than its predecessors. In parallel, xAI's Grok 4.6 and Grok 4.7 joined Bedrock, offering 500K-token context windows, configurable reasoning depth, and self-verification features.
Runtime efficiency and open-source agent tooling
Beyond model catalogs, the September releases targeted agent deployment costs and speed. The updated AgentCore runtime features improved memory management and reduced cold start latency for serverless agents. The infrastructure scales completely to zero during idle periods, operates in hardware-isolated environments, and charges strictly for actual execution rather than reserved peak memory.
On the open-source side, AWS introduced updates to the Strands toolkit. The new Strands harness provides a single-line Python or TypeScript entry point for deploying agents with built-in prompt caching, memory, and context handling, while utilizing 28 percent fewer tokens than comparable harnesses at equivalent accuracy. AWS also published Strands Decider 2B, an open-source 2-billion-parameter model hosted on GitHub and Hugging Face. Designed exclusively to select predefined options rather than generate freeform text, it executes routing and tool-selection decisions locally in roughly 115 milliseconds.
For grounding agents in enterprise information, Amazon Bedrock Managed Knowledge Base introduced native connectors for ServiceNow, Confluence Data Center, Salesforce, and Zendesk. These connectors handle crawling, metadata processing, and incremental updates automatically, complemented by scheduled sync frequencies (daily, weekly, monthly) and user-managed authentication for SharePoint, OneDrive, and Confluence.
What it means for developers
The combined additions give engineering teams granular control over their AI infrastructure budget and design. Instead of routing all tasks through a single heavyweight model, developers can segment workflows: using lightweight options like Strands Decider 2B for local routing, high-throughput engines like GPT 6.1 Luna for data parsing, and high-reasoning tiers like Claude Opus 5.5 or GPT-6 Astra for complex code generation.
Operational overhead is also reduced by serverless AgentCore environments that eliminate the need to pay for idle memory while waiting on asynchronous tool responses. Similarly, native enterprise connectors cut down on bespoke ETL pipelines required to keep knowledge bases synchronized.
As evaluating multiple providers becomes the norm for agentic architectures, teams seeking flexible testing environments can also try top AI models cheaply through one API at https://apixoai.online without managing multiple isolated enterprise contracts. For engineering teams operating inside AWS, the latest Bedrock and Strands updates present a clearer path to matching specific model latency, cost profiles, and governance controls to each production agent task.
Source: ICYMI: What landed for AI builders in September 2026 | Amazon Web Services — AWS Machine Learning. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
