AWS adds models, agent safeguards and cost controls in October

AWS says its October 2026 releases give AI builders more model choices, stronger controls for production agents and clearer ways to manage costs. The updates span Amazon Bedrock, Amazon Bedrock AgentCore and the open source Strands toolkit.
- OpenAI GPT-6.1 Terra is generally available on Amazon Bedrock.
- AgentCore Gateway adds model fallback and support for Step Functions workflows.
- Structured outputs are available for all Claude models on Amazon Bedrock.
- AWS added model-level cost breakdowns to its billing tools at no additional charge.
- Strands Bidi Agents are generally available for speech-to-speech voice agents.
More models for different workloads
AWS added several model options to Amazon Bedrock for workloads ranging from high-volume subagents to real-time voice applications. Claude Haiku 5.5 is designed for subagents and cost-sensitive work. According to Anthropic, it costs around 75 percent less than Claude Haiku 4.5 for most tasks and is the first Haiku model with effort controls.
OpenAI GPT-6.1 Terra is now generally available on Amazon Bedrock. AWS says it performed on par with GPT-6 Astra on OpenAI's Agents' Last Exam evaluation, completed tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, and has input and output token prices at least 50 percent lower than GPT-6.1 Sol and GPT-5.6 Terra. The model accepts text and images with up to 1 million input tokens.
Other additions include Amazon Nova 2.5 Sonic, a speech-to-speech model with lower latency at the same price as Amazon Nova 2 Sonic, plus GLM 5.3 by Z.ai for agentic coding and TwelveLabs Pegasus 1.5 for converting videos into structured, time-coded metadata. AWS also expanded model availability in AWS GovCloud (US), including Claude Haiku 5.5, OpenAI GPT-5.6 models with cross-Region inference and Daybreak Red from OpenAI for authorized security work.
AgentCore targets reliability and workflow control
Amazon Bedrock AgentCore received updates aimed at running agents in production. AgentCore CLI version 1.0 is now generally available, offering a single tool for local development and operations across resources including runtimes, memory, identity, gateways and payments.
AgentCore Gateway now supports Amazon Bedrock InvokeModel APIs and model fallback. If a primary model is throttled or unreachable, it can route requests to another model or provider in a chain defined by the customer. A circuit breaker can skip failing destinations, while the gateway can also reach private targets in a virtual private cloud using certificates from the customer's private certificate authority.
AWS Step Functions workflows can now be used as tools through AgentCore Gateway. This allows an agent to trigger existing Standard and Express workflows while the workflow continues to execute each defined step. Structured outputs are also available for all Claude models on Amazon Bedrock, providing schema-compliant JSON without custom parsing or retries.
AWS further added domain filtering and user location support to Web Search on Amazon Bedrock for requests using the OpenAI Responses API. Amazon Bedrock Guardrails' new Standard tier for contextual grounding checks costs the same as the Classic tier and supports conversational and reasoning responses.
New cost controls and Strands capabilities
AgentCore memory now prices short-term memory according to the data agents ingest, retrieve and store, rather than a flat rate per event. AWS says most workloads will see lower costs, with the largest reductions for small-payload use cases. Long-term memory records can also have a time-to-live, helping control retention and storage costs.
AWS Cost Explorer, AWS Budgets and AWS Cost Management Dashboards now break down Amazon Bedrock spending by model, model provider, inference type and feature at no additional charge. Advanced Prompt Optimization can also shorten prompts when the shorter version scores at least as high in the customer's evaluation.
In Strands, Bidi Agents are generally available for building two-way, speech-to-speech voice agents, with direct support for Amazon Nova Sonic, OpenAI and Gemini models. Strands Box is in developer preview as an open source local sandbox that combines operating-system isolation with semantic policy. Strands Evals can run adversarial conversations, including prompt injections and jailbreak attempts, and can score structured-output fields individually.
For users, the October releases offer more ways to match models to budgets and latency targets while adding fallback, workflow, grounding and output controls for business applications. AWS's cost reporting and memory pricing changes also give finance and platform teams more visibility into the expense of running agents at scale.


