ICYMI：2026 年 9 月 AI 构建者迎来哪些更新
英文标题：ICYMI: What landed for AI builders in September 2026
AWS 在 2026 年 9 月为 Amazon Bedrock、Amazon Bedrock AgentCore 和 Strands 推出了多项更新，覆盖模型选择、Agent 运行时、评估与企业知识连接能力。Amazon Bedrock 新增多款 OpenAI、Claude、Kimi 和 Grok 模型，为不同上下文长度、延迟、吞吐量及成本需求提供更多选择。AgentCore Runtime 通过更高效的内存管理和更低冷启动延迟提升无服务器 Agent 的运行效率，Strands Harness 则宣称在保持准确性的同时可减少 28% 的 Token 使用量。企业还可通过托管知识库的自动同步计划及 ServiceNow、Salesforce、Zendesk 等原生连接器，让 AI Agent 更及时地访问业务信息并降低自定义数据接入维护成本。
1.3 MB
1.3 MB
2.4 KB
345.0 KB
141.2 KB
346.5 KB
25.7 KB
152.5 KB
中文
2026 年 9 月 Amazon Bedrock、Amazon Bedrock AgentCore 和 Strands 最新更新回顾
在 AWS，我们相信，通往企业 AI 的最快路径，是在模型、运行时和工具链等每一层都为构建者提供真正的选择。这一理念塑造了我们对 Amazon Bedrock、AgentCore 和 Strands 的投入方向。Amazon Bedrock 持续扩大对领先前沿模型的访问，让您能够安全地构建和扩展 AI 应用。Amazon Bedrock AgentCore 提供托管基础设施，使您可以使用任何框架或模型来构建、连接和部署企业 Agent。Strands 则提供开源工具包，用于构建 Agent 并将其运行在任何环境中。
随着 AI 模型能力不断提升、模型选择持续扩展，行业讨论的重心正在转移。模型性能已不再是唯一的问题。客户如今会针对特定使用场景权衡成本与收益；当在生产环境中将工作委派给 Agent 时，这些决策尤为重要。AI Agent 已开始进入这样的工作流：它们需要基于企业知识进行推理、处理长期任务，并在安全性、准确性和治理至关重要的系统中采取行动。关注点如今正从模型本身延伸至其周边的更广泛系统，包括 Agent 能够访问的上下文、能够执行的操作，以及企业数据如何得到治理和保护。
9 月，Amazon Bedrock、AgentCore 和 Strands 的更新强化了这些基础能力。新增功能扩展了模型选择，提高了 Agent 运营的效率和可衡量性，并提供了更直接的方式，将 AI 应用连接到最新的业务信息。
运行更快的 Agent，并评估真正重要的事项
**构建针对 OpenAI 模型优化的 Agent。**由 OpenAI 驱动的 Amazon Bedrock Managed Agents 现已进入公开预览阶段，可帮助您在不牺牲企业安全性的前提下构建生产就绪的 AI Agent。现在，您可以在将数据保留在 AWS 内部的同时使用 OpenAI 模型，复用现有的 AWS Identity and Access Management (IAM) 权限，并通过 AWS CloudTrail 保持完整的可审计性。持久会话和内置人工审批工作流有助于降低开发开销并缓解风险。
**Agent 现在可凭借更少的基础设施开销更快启动。**作为 Amazon Bedrock AgentCore 功能的最新版 AgentCore runtime 采用了更高效的内存管理，并为无服务器 Agent 降低了冷启动延迟，因此您只需为实际使用量付费，而非峰值内存。借助按使用量付费的定价模式，您无需预置容量，即可更快、更低成本地运行 Agent。会话在空闲时可缩容至零，并运行在硬件隔离环境中。
在获得前沿性能的同时降低 Token 成本。Strands harness 是一个新的开源 Agent harness，在准确性方面可与主流 harness 媲美，同时 Token 使用量减少 28%。只需一行 Python 或 TypeScript，您就可以启动一个生产就绪的 Agent；它内置上下文管理、提示缓存和记忆功能，随后可部署至任何环境。
让模型在本地安全地做出重大决策。Strands Decider 2B 是一款小型开源 20 亿参数决策模型，可在预定义选项中进行选择，而非生成文本，并能在本地约 115ms 内给出答案。它针对工具选择、路由和护栏等 Agentic AI 任务进行了优化，所有代码、数据和权重均可在 GitHub 和 Hugging Face 上获取。
为每项工作负载找到合适的模型
**为日常工作从更广泛的 OpenAI 模型中进行选择。**OpenAI 的 Astra、Sol 和 Luna 模型现已在 Amazon Bedrock 上全面可用，可分别满足高要求项目、频繁执行的复杂工作和高吞吐量任务的需要。
GPT-6 Astra 是面向最具挑战性工作的旗舰模型，能够为复杂决策、文档分析和软件开发带来更深入的推理能力，并支持最多 100 万个输入 Token。它可为复杂业务决策提供更强的推理能力和更精准的判断，可跨软件和文件开展工作，并生成符合组织语调、模板和标准的专业级输出。GPT-6 Astra Ultrafast 则为 GPT-6 Astra 增加了高级速度层级，面向对速度要求最高的工作负载。其 API 推理速度最高可提升至 6 倍，达到每秒最多 300 个 Token。
Sol 为需要频繁运行的编程、计算机操作和专业工作负载提供接近 Astra 的智能水平。GPT-6.1 Sol 和 GPT-6 Sol 均已在 Amazon Bedrock 上提供。GPT 6.1 Luna 专为提取、摘要、分类和路由等高吞吐量任务设计。它们共同帮助您在不同工作负载之间平衡智能水平、延迟和成本。
**使用 Claude 处理长期运行的编程和研究任务。**Claude 的模型选择现已纳入最新版本，以推动您的编程和科研工作。Claude Fable 5.1 扩展了编程、科学研究和企业工作流的选择范围。Claude Opus 5.5 面向 Agentic 编程、知识工作和长期运行任务而设计。它通过自适应思考判断任务所需的推理程度，并使用 effort 参数设定推理深度的上限。Claude Sonnet 5.5 是更智能、更高效的 Sonnet，适用于聚焦的编程和知识工作；与 Claude Sonnet 5 相比，其单项任务成本降低 30%，速度提高 30%。
**通过更广泛的模型组合应对复杂推理。**Moonshot AI 的 Kimi K3 现已在 Amazon Bedrock 上提供。根据 Moonshot AI 的说法，Kimi K3 是其能力最强的模型，也是首个达到 2.8 万亿参数的开放模型。它拥有 100 万 Token 上下文窗口、原生视觉能力和内置提示缓存，可帮助您以低于先前版本的成本，将智能应用构建速度提升至 2.5 倍。
xAI 的前沿模型 Grok 4.6 和 Grok 4.7 提供 50 万 Token 上下文窗口，并支持可配置的推理强度和自我验证，以提升长期运行任务的可靠性。
这些新增模型共同为您提供了更大的灵活性，可根据上下文长度、延迟、吞吐量和成本等要求匹配相应模型。
获得对企业 AI Agent 的更强控制力
**让 Agent 知识保持最新。**借助 Amazon Bedrock 托管知识库新增的自动同步计划，现在您可以为所有原生数据源连接器配置每日、每周和每月刷新选项。这样一来，您的企业 AI Agent 就能检索最新信息。当您已具备所需访问权限时，针对 SharePoint、OneDrive 和 Confluence 的用户自管设置 可减少对由管理员管理的服务账户的依赖。
**连接更多企业数据源，而无需维护自定义摄取管道。**Amazon Bedrock 托管知识库现已支持 ServiceNow、Confluence Data Center 以及 Salesforce 和 Zendesk 原生连接器，可自动处理数据抓取、元数据提取和增量同步。这些连接器可减少您为构建和维护支持内容、内部文档及运营知识所需的自定义摄取代码量。
开始使用
探索 Amazon Bedrock，使用 AgentCore CLI 部署 Agent，或通过 Strands Harness SDK 构建您的第一个 Agent。
想了解 Amazon Bedrock 如何支持您的团队？请联系我们，开启交流。
关于作者
English / 原文
A recap of the latest Amazon Bedrock, Amazon Bedrock AgentCore, and Strands updates from September 2026
At AWS, we believe the fastest path to enterprise AI is giving builders real choice at every layer: model, runtime, and tooling. That belief shapes how we invest across Amazon Bedrock, AgentCore, and Strands. Amazon Bedrock continues expanding access to leading frontier models so you can build and scale AI applications securely. Amazon Bedrock AgentCore provides the managed infrastructure to build, connect, and deploy enterprise agents using any framework or model. Strands gives you an open source toolkit to build agents and run anywhere.
As AI models become more capable and model choice expands, the industry conversations are shifting. Model performance is no longer the only question. Customers now weigh cost against benefit for their specific use case, and those decisions matter when delegating work to agents in production. AI agents are already moving into workflows that require them to reason over enterprise knowledge, operate across long-horizon tasks, and act in systems where security, accuracy, and governance matter. The focus is now shifting beyond the model itself to the broader system around it, including the context agents can access, the actions they can take, and how enterprise data is governed and protected.
In September, updates across Amazon Bedrock, AgentCore, and Strands strengthened those foundations. New capabilities expanded model choice, made agent operations more efficient and measurable, and created more direct ways to connect AI applications with current business information.
Run faster agents and evaluate what matters
Build agents optimized for OpenAI models. Amazon Bedrock Managed Agents, powered by OpenAI, is now available in public preview, helping you build production-ready AI agents without sacrificing enterprise security. You can now use OpenAI models while keeping data within AWS, reuse existing AWS Identity and Access Management (IAM) permissions, and maintain full auditability through AWS CloudTrail. Durable sessions and built-in human approval workflows help reduce development overhead and mitigate risk.
Agents now start faster with less infrastructure overhead. The latest AgentCore runtime, a capability of Amazon Bedrock AgentCore, has more efficient memory management and lower cold start latency for serverless agents, so you pay for actual usage rather than the peak memory. You can run agents faster at lower cost with pay-as-you-go pricing and no need to pre-provision capacity. Sessions scale to zero when idle and run in hardware-isolated environments.
Lower your token cost with frontier performance. Strands harness is a new open source agent harness that matches popular harnesses on accuracy while using 28 percent fewer tokens. You 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.
Let your models make big decisions safely, locally. Strands Decider 2B is a small, open source, 2B-parameter decision model that picks between predefined options (instead of generating text), delivering answers in ~115ms locally. It’s optimized for agentic AI tasks like tool selection, routing, and guardrails, with all code, data, and weights available on GitHub and Hugging Face.
Find the right model for each workload
Choose from a broader set of OpenAI models for your everyday work. OpenAI’s Astra, Sol, and Luna models are now generally available on Amazon Bedrock, with options for demanding projects, recurring complex work, and high-volume tasks.
GPT-6 Astra is the flagship model for your most ambitious work, bringing deeper reasoning to complex decisions, document analysis, and software development with up to 1 million input tokens. It brings stronger reasoning and more precise judgment to complex business decisions, operates across software and files, and delivers professional-quality output aligned with organizational voice, templates, and standards. GPT-6 Astra Ultrafast adds a premium speed tier for GPT-6 Astra, built for workloads where speed matters most. It delivers up to 6x faster inference in the API, with up to 300 tokens per second.
Sol brings near-Astra intelligence to coding, computer use, and professional workloads that run frequently. Both GPT-6.1 Sol and GPT-6 Sol are available on Amazon Bedrock. GPT 6.1 Luna is designed for high-volume tasks such as extraction, summarization, classification, and routing. Together, they help you balance intelligence, latency, and cost across workloads.
Take on long-running coding and research with Claude. Model choices for Claude now include the latest versions to advance your coding and scientific work. Claude Fable 5.1 expands options for coding, scientific research, and enterprise workflows. Claude Opus 5.5 is built for agentic coding, knowledge work, and long-running tasks. It uses adaptive thinking to gauge how much reasoning a task needs and an effort parameter to set an upper bound on reasoning depth. Claude Sonnet 5.5 is a smarter, more efficient Sonnet suited to focused coding and knowledge work, delivering 30 percent lower cost per task at 30 percent faster speed than Claude Sonnet 5.
Reason through complexity with a broader model portfolio. Moonshot AI’s Kimi K3 is now available on Amazon Bedrock. According to Moonshot AI, Kimi K3 is its 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, it helps you build smarter apps 2.5x faster at lower cost than previous versions.
xAI’s frontier models Grok 4.6 and Grok 4.7 offer a 500K token context window and support configurable reasoning effort and self-verification for more reliable long-running tasks.
Together, these additions give you more flexibility when matching models to your requirements such as context length, latency, throughput, and cost.
Get more control over your enterprise AI agents
Keep agent knowledge current. With new automatic sync scheduling in Amazon Bedrock Managed Knowledge Base, you can now configure daily, weekly, and monthly refresh options for all native data-source connectors. As a result, your enterprise AI agents retrieve the latest information. The user-managed setup for SharePoint, OneDrive, and Confluence reduces reliance on admin-managed service accounts when you already have the access you need.
Connect more enterprise sources without maintaining custom ingestion pipelines. Amazon Bedrock Managed Knowledge Base now supports ServiceNow, Confluence Data Center, Salesforce, and Zendesk as native connectors to handle data crawling, metadata extraction, and incremental sync automatically. These connectors reduce the amount of custom ingestion code you need to build and maintain support content, internal documentation, and operational knowledge.
Get started
Explore Amazon Bedrock, deploy agents with the AgentCore CLI, or build your first agent with the Strands Harness SDK.
Interested in learning how Amazon Bedrock can support your team? Connect with us to start the conversation.