AWS AI Updates: October 14, 2026

1. Grok 4.7 Arrives on Amazon Bedrock With a 500K-Token Context Window

AWS. SpaceXAI’s Grok 4.7 is now available on Amazon Bedrock through cross-Region inference profiles us.xai.grok-4.7 (US-only processing) and global.xai.grok-4.7 (commercial Regions), with text and image input and a 500K-token context window. The model is served through the Bedrock Converse API as well as OpenAI-compatible Chat Completions and Responses APIs, with tool calling, structured output, and configurable reasoning effort from low to xhigh; reasoning tokens count against the output token limit. SpaceXAI-reported results versus Grok 4.6 include Terminal-Bench 4.0 rising from 20.3% to 37.6% and CursorBench 4.0 from 40.4% to 46.3%. In the authors’ own search-task tests, xhigh effort answered 18 of 18 runs correctly using about 9,147 output tokens on average, versus 15 of 18 at low effort with about 1,184 tokens. Source

2. AWS Agent Registry Adds Custom Metadata Schemas for Agents, MCP Servers, and Skills

AWS. AWS Agent Registry, the catalog for an organization’s agents, MCP servers, agent skills, gateways, and custom resources, now supports custom metadata on registry records. Administrators define a schema once per registry and can extend it with type-specific fields, such as a model family for agent records, using free text, URL, Boolean, or enumerated values that the registry validates on every create or update. Search results can be filtered on any custom field from the console, CLI, SDKs, or the registry’s MCP endpoint. The feature is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Sydney), Asia Pacific (Tokyo), and Europe (Ireland). Source

3. AWS AI/ML Skill Becomes a Core Skill in the Agent Toolkit for AWS

AWS. The AWS AI/ML skill, which lets coding agents select, train, optimize, and deploy models on Amazon SageMaker AI, is now a core skill in the Agent Toolkit for AWS, which AWS says replaces the earlier AWS Labs skills and plugins as the single supported integration point. The skill’s coverage now spans inference optimization for latency, throughput, and cost, guided fine-tuning from use case definition through deployment, serverful training, MLflow experiment tracking, JumpStart model deployment and recommendations, and processing jobs. It comes pre-installed in SageMaker Studio JupyterLab for use with Kiro or other coding agents and carries no additional charge beyond the resources agents consume. Source

4. AWS Upstreams MoE Kernel Optimizations to vLLM, Reports Up to 10% Throughput Gain on H200

AWS. An AWS post details three fused Mixture-of-Experts inference optimizations for NVIDIA H200 GPUs, now upstreamed to vLLM and shipped in the AWS LMI Deep Learning Container for SageMaker AI. A “swap-AB” transpose places the large expert dimension in the Hopper WGMMA operand-A slot to fix low Tensor Core utilization during decode, yielding 2% to 4% higher end-to-end throughput (vLLM 0.24.0 and later). Routing FP8 tensor-quantized MoE models to the Triton backend instead of FLASHINFER_CUTLASS gave a 10% throughput gain with no kernel changes (LMI v28), and adding do_not_specialize annotations to stop per-batch-size Triton JIT recompilation cut time to first token by 9% at 8-way tensor parallelism (vLLM 0.22.0 and later). Source

5. GPT-6 Models on Bedrock Get Global Cross-Region Inference From Five Southeast Asian Regions

AWS. OpenAI’s GPT-6 Sol and GPT-6 Luna on Amazon Bedrock can now be called through global cross-Region inference from Singapore, Jakarta, Malaysia, Thailand, and Taipei, while GPT-6 Astra is available from Singapore only, using profile IDs such as global.openai.gpt-6-sol. Billing, quota tracking, CloudTrail, and CloudWatch stay in the source Region, and the endpoint supports the Converse API plus OpenAI Responses and Chat Completions paths, including configuration for Codex CLI. Quotas are measured in tokens per minute, with each output token consuming 10 units of quota and the full max_tokens reserved up front. AWS notes that Region-allowlist SCPs can block these calls and that traffic flagged by abuse classifiers may be retained for up to 30 days in the destination Region. Source

6. AWS Publishes Automation for Activating Bedrock Private Pricing Offers Across Organizations

AWS. A new AWS post shows how to automate activation of AWS Marketplace private pricing offers for third-party models on Amazon Bedrock using managed entitlements in AWS License Manager. Organization-level grants created after a buyer accepts an offer are auto-accepted but stay disabled, and negotiated pricing does not apply until a distributed grant is activated for the organization, selected OUs, or specific accounts. The reference design uses an EventBridge rule in us-east-1 on Marketplace agreement events, a Lambda function that checks a DynamoDB seller allow-list before creating and activating grants, and a CDK v2 deployment, with a lighter dry-run script for one-time use. Source