AWS's machine learning blog published its October 2026 recap for AI builders, covering Amazon Bedrock, Bedrock AgentCore and the open source Strands toolkit. It is a vendor summary, so the claims below are AWS's own or attributed to the model makers.
New models
- Claude Haiku 5.5 targets high-volume subagent work. Anthropic claims that, for the majority of tasks, its model runs at roughly a 75 percent lower cost than Claude Haiku 4.5. It is also the first Haiku release to include effort controls, which let users balance expense against capability. AWS suggests a larger model such as Claude Opus 5.5 plan the work and pass defined tasks to Haiku.
- OpenAI GPT-6.1 Terra is generally available on Bedrock. AWS reports it matched GPT-6 Astra on OpenAI's Agents' Last Exam while finishing tasks more than three times faster, with token prices at least 50 percent below GPT-6.1 Sol and GPT-5.6 Terra. It takes text and images, up to 1 million input tokens.
- Amazon Nova 2.5 Sonic, a speech-to-speech model, improves reasoning and tool-calling accuracy with lower latency, at the same price as Nova 2 Sonic.
- GLM 5.3 (open-weight, for coding) and TwelveLabs Pegasus 1.5 (video to time-coded metadata) were also added. Several models, including Claude Haiku 5.5, reached AWS GovCloud (US).
Reliability and control
AgentCore Gateway now supports the Bedrock InvokeModel APIs and model fallback. When the primary model is throttled or unreachable, requests move along a chain of models or providers that you define, and a circuit breaker skips failing destinations. Existing Bedrock SDK code can point at the gateway without per-provider changes.
Standard and Express Step Functions workflows can now be used as agent tools through the same gateway. The workflow keeps its approval and eligibility steps, and the agent only chooses when to start it. Structured outputs, which return schema-compliant JSON, now cover all Claude models on Bedrock. Web Search supports allow and block lists of up to 100 domains each, and Guardrails has a Standard tier for grounding checks at the Classic tier's price.
Cost
AgentCore short-term memory is now priced by data ingested, retrieved and stored. AWS says most workloads will pay less, with small-payload cases saving the most. Long-term memory records can carry a time-to-live. Cost Explorer, Budgets and Cost Management Dashboards now break Bedrock spend down by model, provider, inference type and feature at no extra charge.
Strands
Strands Bidi Agents for voice are generally available. Strands Box, a local sandbox with per-tool policy, is in developer preview. Strands Evals adds red teaming and field-level scoring of structured output. Strands Decider, a small model that picks between options you define, now comes in four sizes. AWS says one internal ticket-classification system saw cost per ticket fall by about 98 percent after switching to it. That figure is AWS's own estimate and has not been independently verified.
Top comments (0)