Amazon Bedrock adds model fallback, Claude JSON outputs

AWS frames the October 2026 release around three practical needs for teams running AI agents: choosing a model that fits each task's budget and latency, keeping agents working when a model misbehaves, and seeing what the workload costs. On the model side, Claude Haiku 5.5 is positioned for high-volume subagents and, according to Anthropic, costs around 75 percent less than Claude Haiku 4.5 for most tasks. OpenAI GPT-6.1 Terra is now generally available on Amazon Bedrock, with input and output token prices at least 50 percent lower than GPT-6.1 Sol and GPT-5.6 Terra. Amazon Nova 2.5 Sonic, the latest speech-to-speech model, improves reasoning and tool-calling accuracy at the same price as Amazon Nova 2 Sonic.

The reliability updates center on AgentCore Gateway, which now supports model fallback and a circuit breaker so agents can keep working when a primary model is throttled or unreachable. Structured outputs for all Claude models on Amazon Bedrock mean builders receive schema-compliant JSON, which AWS says removes custom parsing and retries. Web Search on Amazon Bedrock gained domain filtering and user location, and Amazon Bedrock Guardrails added a Standard tier for contextual grounding checks at the same cost as the Classic tier.

On cost, AgentCore memory now bills short-term memory by usage rather than a flat rate per event, and AWS Cost Explorer, AWS Budgets, and AWS Cost Management Dashboards break down Amazon Bedrock spend by model, provider, inference type, and feature at no additional charge. The Strands toolkit also expanded: Bidi Agents for voice are generally available, Strands Box is in developer preview as a local sandbox, and Strands Decider now comes in four sizes built on Gemma 4. An Amazon team that switched ticket classification to Decider saw estimated cost per ticket drop by about 98 percent and average response time drop by 78 percent.