AWS AI Updates: October 15, 2026

1. SageMaker Studio Adds Google, GitHub, and Apple Sign-Up With Free Credits

AWS. New customers can now sign up for Amazon SageMaker Studio with Google, GitHub, or Apple credentials and land in a preconfigured environment without creating IAM configuration or navigating the AWS console. Sign-up includes AWS Free Tier credits with no credit card required, and verified customers get default access to G5, G6, G6e, and G4dn instances for endpoints, training jobs, and notebooks without filing quota increase requests. Teams can invite collaborators by email to share models, notebooks, and endpoints, and accounts that exceed free credits can move to a paid plan with a spend limit. The flow is also reachable from the “Deploy on SageMaker AI” buttons on supported Hugging Face model pages, which lowers the setup cost of trying fine-tuning and deployment on AWS. Source

2. SageMaker JumpStart Walkthrough Covers New Model Hub and Pre-Configured Endpoint Profiles

AWS. An AWS walkthrough shows a discover, deploy, test, and clean-up loop for JumpStart models using SageMaker Python SDK v3 (ModelBuilder.from_jumpstart_config), and highlights two 2026 additions: a Model Hub in the SageMaker AI console for browsing the catalog and opening sample notebooks in Studio, and named pre-configured endpoint options tuned for low cost, low latency, or general-purpose workloads, including Neuron-optimized configurations. The sample script records deployment time, request latency, and words per second; a single run of a Qwen 27B vision-language model returned 153 words in 9.07 seconds. The post includes no quality benchmark, so the script is a latency and cost smoke test rather than an evaluation harness. Source

3. AgentCore Sample Builds Document Extraction That Learns Per-Vendor Instructions

AWS. AWS published an open-source guidance sample with eight Strands agents hosted on Bedrock AgentCore Runtime that extract data from documents and store refined extraction instructions for reuse. Unseen document types go through a training graph where a Validator checks output against schemas and business rules in Aurora DSQL and an Instructions Fixer rewrites prompts on failure; later documents are matched through S3 Vectors and routed to an automation path when the similarity score clears a threshold. AgentCore Gateway exposes Lambda validators as MCP tools, and the stack uses Amazon Nova and Claude models on Bedrock. In AWS’s example a vendor’s second order took 1 minute 30 seconds versus 3 minutes 49 seconds for the first, and AWS estimates about $0.055 per document at 1,000 documents per month, with roughly 85% of cost from inference. Source

4. AWS Sample Generates PowerPoint Decks From Long Documents on AgentCore Runtime

AWS. A new aws-samples repository builds a supervisor agent on Bedrock AgentCore Runtime that turns long source documents into .pptx files with speaker notes, grounded through Bedrock Knowledge Bases on OpenSearch Serverless. It uses Claude Sonnet 4.6 for reasoning and content, Claude Haiku 4.5 for document parsing, Titan Text Embeddings v2, and Stability AI Stable Image Inpaint for slide images, with a containerized Lambda generating the PPTX and Bedrock Guardrails applied. AWS reports about 70 to 90 seconds and roughly $0.04 in model costs for a six-slide deck from a 90-page annual report. The sample is deployable with AWS CDK and fronted by a .NET 8 API. Source