AWS AI Updates: October 13, 2026

1. Bedrock AgentCore Moves Short-Term Memory to Consumption-Based Pricing

AWS. Amazon Bedrock AgentCore now bills short-term memory (STM) by the data agents ingest, retrieve, and store, replacing the previous flat rate per event. The new rates are $1.00 per GB for ingestion, $0.20 per GB for retrieval, and $0.10 per GB-month for storage, with ingestion and retrieval billed per event at a 12 KB minimum and 64 KB maximum billable size. AWS says most workloads will see lower costs, with the largest reductions for agents that write frequent, small events, and storage costs can be trimmed by setting a shorter TTL on memory resources. Long-term memory pricing is unchanged, and the new model applies in all Regions where AgentCore Memory is offered. Source

2. Ray LLM Deep Learning Container Adds Multi-Node Serving With Prefill/Decode Disaggregation

AWS. AWS launched a Ray LLM Deep Learning Container for multi-node deployments that serves a model across several nodes behind one OpenAI-compatible endpoint. The image bundles vLLM, Ray Serve, NIXL for KV-cache transfer between nodes, and the EFA networking stack (libfabric and aws-ofi-nccl). The walkthrough runs Qwen3.5-9B in bfloat16 on Amazon EKS with KubeRay, splitting prefill and decode across two g6.8xlarge GPU nodes in a single Availability Zone, and reports model-weight loading on restart dropping from about 145 seconds to about 2 seconds with the described cache configuration. The post does not publish throughput or latency benchmarks. Source

3. AWS Shows How to Serve Decision Models With the llama.cpp Container on SageMaker AI

AWS. A new AWS post details deploying decision models, which return probabilities over a fixed set of application-defined options instead of generating text, using the AWS llama.cpp Deep Learning Container on Amazon SageMaker AI. The setup extends the container with a SemIf (formerly OpenJev) adapter built in AWS CodeBuild and serves Qwen3.5-4B in Q4_K_M quantization. On 231 JevBench tasks, an ml.g6e.xlarge GPU endpoint reached 77.9% accuracy at 0.352 s p50 and 0.599 s p95 latency, versus 77.5% at 1.715 s p50 and 19.291 s p95 on an ml.c6i.8xlarge CPU instance; smaller 2B and 0.8B models cut latency but dropped to 66.7% and 58.4% accuracy. The post gives step-by-step code for routing, prioritization, and agent action selection use cases. Source

4. Amazon Connect Customer Prefills After-Contact Guides From Conversation Data

AWS. Amazon Connect Customer can now populate the step-by-step guides that contact center agents complete after a call, using information extracted from the conversation such as booking identifiers, claim numbers, and conversation summaries. Managers define which extracted values map to each form field, and agents review, correct, and submit instead of typing each field manually. The feature is available in nine Regions including US East (N. Virginia), US West (Oregon), Europe (Frankfurt), and Asia Pacific (Tokyo), with contact summaries also available in Africa (Cape Town). Source