AWS Expands Bedrock and Strands with New Model Options and Production Tools

For voice interactions, AWS launched Amazon Nova 2.5 Sonic, a speech-to-speech model offering lower latency, improved reasoning, and stronger tool execution at the same price as Nova 2 Sonic. For specialized workloads, Z.ai's GLM 5.3 brings a 1-million-token context window with prompt caching for enterprise software engineering, while TwelveLabs Pegasus 1.5 extracts structured, time-stamped metadata from video, audio, and on-screen text in a single pass. In AWS GovCloud (US), regulated customers now have access to Claude Haiku 5.5, GPT-5.6 models with cross-Region inference, and OpenAI's Daybreak Red for authorized security operations.

AgentCore tooling and enterprise safeguards

To manage agents beyond the prototype stage, AWS released AgentCore CLI 1.0, consolidating local development and operations for runtimes, memory, identity, gateways, and payments under a unified command structure.

AgentCore Gateway now incorporates automated model fallback and circuit breakers for Bedrock InvokeModel APIs. If a primary model becomes unreachable or throttled, traffic routes to a designated backup provider without requiring custom client-side retry logic. The gateway also connects to private targets inside virtual private clouds (VPCs) via private certificate authorities, and allows teams to invoke AWS Step Functions workflows as standard agent tools.

Structured outputs are now enabled across all Claude models on Bedrock, returning schema-validated JSON without secondary extraction code. For information retrieval, Bedrock Web Search via the OpenAI Responses API now allows developers to allowlist or blocklist up to 100 domains and filter results by location, while Bedrock Guardrails added a Standard tier for contextual grounding checks across conversational and reasoning responses.

AWS also revised its billing mechanisms. AgentCore short-term memory shifted to usage-based pricing based on data ingested, retrieved, and stored rather than a flat per-event fee. In AWS Cost Explorer, Budgets, and Cost Management Dashboards, teams can now break down Bedrock expenditures by model, provider, feature, and inference type.

Local isolation, testing, and routing in Strands

The open-source Strands framework received updates focused on agent safety, voice execution, and deterministic routing. Strands Box entered developer preview as an open-source sandbox combining operating-system isolation with semantic policies. It restricts an agent's access to files and external services regardless of whether tools are accessed directly, via generated code, or through the Model Context Protocol (MCP).

Strands Bidi Agents reached general availability, supporting continuous two-way audio streams with Nova Sonic, OpenAI, and Gemini models by handling reconnections behind the scenes when session limits expire. For testing, Strands Evals added red-teaming simulations for prompt injections and jailbreaks, alongside field-by-field scoring for structured outputs.

Finally, the framework's small routing model, Strands Decider—built on Gemma 4 across four sizes—offers an alternative to generative text models by making discrete operational choices. AWS reported that migrating an internal ticketing system to Decider reduced estimated cost per ticket by approximately 98 percent and cut response times by 78 percent.

What it means for developers

These updates show AWS addressing the operational pain points of running multi-agent systems at scale: high token costs, throttling risks, and execution safety. By adding automated fallback inside AgentCore Gateway and native schema enforcement on Claude models, teams spend less time engineering defensive wrapper code and retry systems.

Cost management is also shifting from guesswork to granular controls. With subagent-tier models like Haiku 5.5, specialized deciders like Gemma-based Strands Decider, and prompt-shortening capabilities in Advanced Prompt Optimization, developers can systematically route non-critical steps away from large frontier models.

For teams working across multiple foundation model providers, orchestration flexibility remains essential. Developers can also try top AI models cheaply through one API at https://apixoai.online to test and compare different architectures. Paired with local sandboxing tools like Strands Box, builders have clearer paths to move autonomous agents from experimental scripts into governed production environments.

Source: ICYMI: What landed for AI builders in October 2026 | Amazon Web Services — AWS Machine Learning. Written by the Apixo team from that report.