11-19 更新 实时
首页/ 算力/ 正文

ICYMI:2026 年 10 月 AI 开发者的重要更新

算力 来源:AWS Machine Learning 2026-11-12 06:26 0 阅读 机器翻译
ICYMI:2026 年 10 月 AI 开发者的重要更新

在 AWS,我们相信 AI 构建者应当能够为每项任务选择正确的模型、在生产环境中信任自己的智能体,并清晰掌握其 AI 工作负载的成本。这一信念指引着我们在整个产品组合中的投资方向。Amazon Bedrock 让你能够获取领先模型,以及安全构建和扩展 AI 应用的工具。我们通过 Amazon Bedrock AgentCore 进一步扩展这一基础,它提供了用任意模型和任意框架构建、连接和优化智能体的平台。这其中包括 Strands——我们的开源智能体开发工具包。二者结合,带你从第一个原型走向生产环境。

随着智能体进入核心业务流程,门槛也随之提高。当任务、预算和延迟目标发生变化时,你需要灵活使用不同的模型。当主模型被限流或无法访问时,你的智能体必须继续工作。它们还必须在你能设定的限制内行动,并返回其他软件可直接使用的输出。它们的答案必须始终基于你允许的来源,而且其成本在大规模使用下也必须撑得住。

我们在 10 月发布的内容回应了上述每一项需求。新模型帮助你为每项任务匹配合适的价格与速度,从高并发的子智能体到实时语音。Amazon Bedrock 和 AgentCore 的更新让智能体保持可用、使其遵循你定义的流程,并帮助你信任它们返回的结果。新的定价与成本报告让支出更可预测。Strands 新增了构建、管控和测试智能体的方式。

运行快速、低成本的子智能体。Claude Haiku 5.5 专为子智能体和高频、成本敏感型工作打造。据 Anthropic 称,它是迄今能力最强的 Haiku,在大多数任务上的成本比 Claude Haiku 4.5 低约 75%。它也是首个支持力度控制的 Haiku,因此你可以针对每项任务在成本与智能之间进行权衡。在多智能体架构中,Claude Opus 5.5 这类模型可以负责规划工作,并把定义清晰的编码、工具调用和浏览器自动化任务交给 Haiku。这样一来,并行运行大量智能体依然成本可控。

以更低价格获得接近旗舰级的效果。OpenAI GPT-6.1 Terra 现已在 Amazon Bedrock 上正式可用。在 OpenAI 的 Agents' Last Exam 评测中,它的表现与 GPT-6 Astra 相当,完成任务的速度是 GPT-6 Astra 和 GPT-6.1 Sol 的三倍以上。其输入和输出 Token 价格比 GPT-6.1 Sol 和 GPT-5.6 Terra 至少低 50%。它可接受文本和图像,输入最多支持 100 万 Token,适合起草提案、解读技术图纸,以及构建跨业务系统工作的智能体。对于实时编程助手等对延迟敏感的工作,GPT-6.1 Sol 现在还提供 Ultrafast 模式。

构建可实时推理的语音智能体。我们最新的语音到语音模型 Amazon Nova 2.5 Sonic 提升了推理、指令遵循和工具调用准确性,在与 Amazon Nova 2 Sonic 价格相同的情况下实现了更低的延迟。客服智能体现在可以在一次对话中查询订单、核对退货资格并发起退货。

为编码和视频挑选专用模型。Z.ai 的 GLM 5.3 是一款面向智能体式编码和长周期软件工程的开放权重模型,具备 100 万 Token 上下文窗口和提示缓存,面向符合条件的企业客户开放。TwelveLabs Pegasus 1.5 可将整段视频转化为结构化、带时间码的元数据。它一次即可读取画面、音频和屏幕文字,无需单独的数据摄取流程。

在 AWS GovCloud (US) 中使用更多模型。支持跨区域推理的 Claude Haiku 5.5 和 OpenAI GPT-5.6 模型现已在 AWS GovCloud (US) 上线。OpenAI 的 Daybreak Red——一款用于授权安全工作的模型——已面向 AWS GovCloud (US-East) 的符合条件客户提供。受监管团队在保持美国数据主权的同时,可以使用更多模型。

在终端中构建和运行智能体。作为 Amazon Bedrock AgentCore 的一项能力,AgentCore CLI 现已正式发布 1.0 版本。我们对其进行了重构,启动更快、命令结构统一,并支持更多 AgentCore 资源。你可以直接或通过项目来使用 harness、运行时、记忆、身份、网关和支付。现有项目只需一条命令即可迁移到 1.0 版本。最终结果是:从本地开发到日常运维只需一个工具。

当模型被限流时,让智能体保持可用。AgentCore Gateway 是 Amazon Bedrock AgentCore 的一项能力,现已支持 Amazon Bedrock InvokeModel API 和模型回退。你现有的 Amazon Bedrock SDK 代码可以直接指向该网关,无需针对不同提供商修改代码。当主模型被限流或无法访问时,网关会按你定义的链路,将请求路由到下一个模型或提供商。断路器会跳过出现故障的目标。该网关还可以使用你自己的私有证书颁发机构颁发的证书,访问你虚拟私有云(VPC)中的私有目标。你的智能体可以保持可用,而无需在每个客户端中编写重试逻辑。

将你的业务工作流转为智能体工具。现在,你可以通过 AgentCore Gateway 将 AWS Step Functions 工作流(包括你现有的 Standard 和 Express 工作流)作为工具使用。你的业务逻辑仍保留在你已经在运行且信任的工作流中,因此无需在智能体内重建,也不用依赖提示词来遵循它。由智能体决定何时运行工作流,而 Step Functions 会严格按照你的定义执行每一个步骤。例如,客服智能体可以启动你的退款工作流,而每笔退款仍会经过相同的资格校验、审批和入账步骤。

获得你的系统可直接使用的 JSON。结构化输出现已在 Amazon Bedrock 上对所有 Claude 模型可用。你可以获得符合 schema 的 JSON,无需自定义解析或重试。你可以从财报中提取所有必需字段,或对支持工单进行路由,其中的实体始终与你的数据模型匹配。

让答案立足于你信任的来源。Amazon Bedrock 上的 Web Search 现在支持针对通过 OpenAI Responses API 发起的请求进行域名过滤和用户位置设置。你可以允许最多 100 个域名、屏蔽最多 100 个域名,并返回与某一地理区域相关的结果。随后由 Amazon Bedrock Guardrails 检查答案是否忠于其来源。其用于上下文溯源校验的新 Standard 层级与 Classic 层级价格相同,覆盖对话式与推理式回答,并支持更多语言。两者结合,你既能控制答案的来源,也能在更多应用中捕获幻觉。

按实际用量为智能体记忆付费。AgentCore memory 是 Amazon Bedrock AgentCore 的一项能力,其短期记忆现在按你的智能体摄入、检索和存储的数据计费,而不再采用按事件统一收费的方式。大多数工作负载的成本会下降,小数据量用例的降幅最大。你还可以为长期记忆记录设置存活时间(TTL),以满足留存要求并控制存储成本。

看清是什么推高了你的 Amazon Bedrock 支出。AWS Cost Explorer、AWS Budgets 和 AWS Cost Management Dashboards 现在可按模型、模型提供商、推理类型和功能拆分 Amazon Bedrock 成本,且不额外收费。财务和平台团队可以在他们已经使用的账单工具中,找出成本最高的模型,或为单个模型设置预算提醒。为降低输入成本,Advanced Prompt Optimization 现在还能精简你的提示词,仅当更短版本在你的评估中得分不低于原版本时才会返回该版本。

只需几行代码即可交付语音智能体。用于构建双向、语音到语音语音智能体的 Strands Bidi Agents 现已正式可用,并直接支持 Amazon Nova Sonic、OpenAI 和 Gemini 模型。Bidi Agents 会在后台重新连接模型,因此长时间通话可以突破提供商的会话时长上限继续下去。语音智能体还可以把更耗时长的工作(例如调研)交给另一个智能体,并朗读解释结果。

让本地智能体在你的限制范围内工作。现已进入开发者预览的 Strands Box 是一个面向 AI 智能体的开源本地沙箱。它把操作系统隔离与语义策略结合起来。你为每个工具设定策略,Box 会确定性地执行该策略——无论智能体是直接调用工具、通过 Model Context Protocol(MCP)调用,还是通过生成代码调用。Box 可与任何模型和任何智能体框架配合使用,因此智能体可以在你的项目中运行,而不会触及你未允许的文件和服务。

在客户发现之前,先找出你的智能体在哪里出问题。Strands Evals 中的红队测试会针对你的智能体运行对抗性对话,例如提示注入和越狱尝试,并展示其失败处的消息和工具调用。Strands Evals 现在还能逐字段对结构化输出打分,当智能体大多数字段都答对时给出部分得分。用它们在生产环境上线之前修复薄弱环节。要开始使用,请参阅 Strands Evals 文档。

原文较长,译文为节选;译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(ICYMI: What landed for AI builders in October 2026)

At AWS, we believe AI builders should be able to choose the right model for every task, trust their agents in production, and see clearly what their AI workloads cost. That belief guides how we invest across our portfolio. Amazon Bedrock gives you access to leading models and the tools to build and scale AI applications securely. We extend this foundation with Amazon Bedrock AgentCore, which provides the platform to build, connect, and optimize agents using any model and any framework. That includes Strands, our open source agent development toolkit. Together, they take you from first prototype to production.

As agents move into core business workflows, the bar rises. You need the flexibility to use different models as tasks, budgets, and latency targets change. When a primary model is throttled or unreachable, your agents must keep working. They also must act within limits you set and return output other software can use as is. Their answers must stay grounded in sources you allow, and their cost must hold up at high volume.

What we shipped in October addresses each of these needs. New models help you match each task to the right price and speed, from high-volume subagents to real-time voice. Amazon Bedrock and AgentCore updates keep agents available, hold them to the processes you define, and help you trust what they return. New pricing and cost reporting make spend more predictable. Strands adds new ways to build, contain, and test agents.

Match the model to the task on Amazon Bedrock

Run fast, low-cost subagents. Claude Haiku 5.5 is built for subagents and high-volume, cost-sensitive work. According to Anthropic, it is the most capable Haiku yet and costs around 75 percent less than Claude Haiku 4.5 for most tasks. It is also the first Haiku with effort controls, so you can tune cost against intelligence for each task. In a multi-agent setup, a model such as Claude Opus 5.5 can plan the work and hand well-defined coding, tool use, and browser automation tasks to Haiku. This keeps running many agents in parallel affordable.

Get near-flagship results at a lower price. OpenAI GPT-6.1 Terra is now generally available on Amazon Bedrock. On OpenAI’s Agents’ Last Exam evaluation, it performed on par with GPT-6 Astra and completed tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. Its input and output token prices are at least 50 percent lower than those of GPT-6.1 Sol and GPT-5.6 Terra. It accepts text and images with up to 1 million input tokens, which suits drafting proposals, interpreting technical drawings, and building agents that work across business systems. For latency-sensitive work such as real-time coding assistants, GPT-6.1 Sol now also offers Ultrafast mode.

Build voice agents that reason in real time. Amazon Nova 2.5 Sonic, our latest speech-to-speech model, improves reasoning, instruction following, and tool-calling accuracy, with lower latency at the same price as Amazon Nova 2 Sonic. A customer service agent can now look up an order, check return eligibility, and start a return in a single conversation.

Pick specialized models for coding and video. GLM 5.3 by Z.ai is an open-weight model for agentic coding and long-horizon software engineering, with a 1-million-token context window and prompt caching, available to eligible enterprise customers. TwelveLabs Pegasus 1.5 turns entire videos into structured, time-coded metadata. It reads visuals, audio, and on-screen text in a single pass, with no separate ingestion pipeline.

Use more models in AWS GovCloud (US). Claude Haiku 5.5 and OpenAI GPT-5.6 models with cross-Region inference are now available in AWS GovCloud (US). Daybreak Red from OpenAI, a model for authorized security work, is available to eligible customers in AWS GovCloud (US-East). Regulated teams get more models while maintaining US data sovereignty.

Build and run agents you can rely on

Build and operate agents from your terminal. AgentCore CLI, a capability of Amazon Bedrock AgentCore, is now generally available in version 1.0. We rebuilt it with faster startup, a consistent command structure, and support for more AgentCore resources. You can work with harnesses, runtimes, memory, identity, gateways, and payments directly or through a project. Existing projects move to version 1 with one command. The result is one tool from local development to day-to-day operations.

Keep agents available when a model is throttled. AgentCore Gateway, a capability of Amazon Bedrock AgentCore, now supports the Amazon Bedrock InvokeModel APIs and model fallback. Your existing Amazon Bedrock SDK code can point at the gateway without per-provider code changes. When the primary model is throttled or unreachable, the gateway routes the request to the next model or provider in a chain you define. A circuit breaker skips destinations that are failing. The gateway can also reach private targets in your virtual private cloud (VPC) using certificates from your own private certificate authority. Your agents stay available without retry logic in every client.

Turn your business workflows into agent tools. You can now use AWS Step Functions workflows, including your existing Standard and Express workflows, as tools through AgentCore Gateway. Your business logic stays in a workflow you already run and trust, so you don’t rebuild it inside the agent or rely on a prompt to follow it. The agent decides when to run a workflow, and Step Functions runs each step exactly as you defined it. For example, a support agent can start your refund workflow, and every refund still goes through the same eligibility check, approval, and credit steps.

Get JSON your systems can use as is. Structured outputs are now available for all Claude models on Amazon Bedrock. You receive schema-compliant JSON without custom parsing or retries. You can extract every required field from an earnings report, or route support tickets with entities that always match your data model.

Keep answers grounded in sources you trust. Web Search on Amazon Bedrock now supports domain filtering and user location for requests made with the OpenAI Responses API. You can allow up to 100 domains, block up to 100 domains, and return results relevant to a geography. Amazon Bedrock Guardrails then checks that answers stay true to their sources. Its new Standard tier for contextual grounding checks costs the same as the Classic tier, covers conversational and reasoning responses, and supports more languages. With both, you can control where answers come from and catch hallucinations in more of your applications.

Predict and lower the cost of running agents

Pay for agent memory based on what you use. AgentCore memory, a capability of Amazon Bedrock AgentCore, now prices short-term memory by the data your agents ingest, retrieve, and store, instead of a flat rate per event. Most workloads see lower costs, and small-payload use cases see the largest reductions. You can also set a time-to-live on long-term memory records to meet retention requirements and control storage costs.

See what drives your Amazon Bedrock spend. AWS Cost Explorer, AWS Budgets, and AWS Cost Management Dashboards now break down Amazon Bedrock cost by model, model provider, inference type, and feature, at no additional charge. Finance and platform teams can find their most expensive models or set a budget alert for a single model, in the billing tools they already use. To lower input costs, Advanced Prompt Optimization can now also shorten your prompts, returning a shorter version only when it scores at least as high on your evaluation.

Build voice, local, and well-tested agents with Strands

Ship voice agents in a few lines of code. Strands Bidi Agents, for building two-way, speech-to-speech voice agents, are now generally available, with direct support for Amazon Nova Sonic, OpenAI, and Gemini models. Bidi Agents reconnect to the model in the background, so long calls keep going past a provider’s session limit. A voice agent can also hand longer work, such as research, to another agent and explain the results aloud.

Let local agents work within your limits. Strands Box, now in developer preview, is an open source local sandbox for AI agents. It combines operating-system isolation with semantic policy. You set policy for each tool, and Box enforces it deterministically, whether the agent calls tools directly, through the Model Context Protocol (MCP), or by generating code. Box works with any model and any agent framework, so an agent can operate in your project without reaching files and services you didn’t allow.

Find where your agent breaks before customers do. Red teaming in Strands Evals runs adversarial conversations, such as prompt injections and jailbreak attempts, against your agent and shows the messages and tool calls where it failed. Strands Evals can now also score structured output field by field, giving partial credit when an agent gets most fields right. Use them to fix weak spots before an agent reaches production. To get started, see the Strands Evals documentation.

Cut the cost of routine agent decisions. Strands Decider, the small, open source decision model we featured in last month’s recap, now comes in four sizes built on Gemma 4. Instead of generating text, it picks between options you define, such as which tool to call or where to route a request. When an Amazon team switched ticket classification in one of its largest internal ticketing systems to Decider, estimated cost per ticket dropped by about 98 percent. Average response time dropped by 78 percent.

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.

 


About the author

Tymofii Kryvtsun

Tymofii Kryvtsun

Tymofii Kryvtsun is a Technical Product Marketing Manager for the AI portfolio at Amazon Web Services (AWS), where he helps developers and enterprises worldwide build production-ready AI agents.

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 AWS Machine Learning(ICYMI: What landed for AI builders in October 2026),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗