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.
