AWS AI Updates: October 21, 2026

1. Bedrock Guardrails Adds a Standard Tier for Contextual Grounding Checks

AWS. Amazon Bedrock Guardrails now offers a Standard tier for contextual grounding checks, which flag responses that are not grounded in a reference source or not relevant to the user’s query. Unlike the Classic tier, which applied strict grounding suited to summarization, paraphrasing, and question answering, the Standard tier handles responses that include reasoning or conversational content, scores differently worded responses with the same meaning more consistently, and supports more languages. The tier is chosen per guardrail, existing guardrails stay on Classic until changed, pricing is the same for both tiers, and the Standard tier requires cross-Region inference with a guardrail profile. It is available in 11 Regions across the US, Europe, and Tokyo. Source

2. Bedrock Advanced Prompt Optimization Can Now Shorten Prompts

AWS. Amazon Bedrock Advanced Prompt Optimization adds an optional prompt shortening setting that reduces prompt length while maintaining or improving scores on the user’s own evaluation criteria. The optimizer returns the shortened prompt only if it is both shorter than the original and scores at least as high; otherwise the original prompt is returned. Results include a compression ratio and the original and optimized token counts, and the setting works with steering criteria, LLM-as-a-judge rubrics, and custom Lambda evaluators. It is available in all Regions where Advanced Prompt Optimization is offered. Source

3. Amazon Polly Generative Engine Gets a New Model, 14 Voices, and Hindi and Finnish

AWS. Amazon Polly’s Generative engine now runs on a new model that AWS says produces more natural and accurate speech with low latency, including more reliable reading of numbers, email addresses, and street addresses. The drop-in upgrade adds 14 voices across nine locales and makes Hindi and Finnish generally available on the Generative engine for the first time, while SSML control continues to work as before. AWS positions the low-latency streaming API for contact centers and LLM-based voice agents, and the launch also extends Polly Generative to AWS GovCloud (US-West). Source

4. i14 Trains and Serves the Laplace 3B Linear Attention Model on Trainium2 With TorchNeuron

AWS. AI research company i14 and the AWS Generative AI Innovation Center brought Laplace 3B, a Gated Linear Attention model whose recurrent state stays fixed at 29.5 MB per sequence regardless of context length, to Trainium2 using the TorchNeuron native PyTorch backend. The team replaced GPU-specific Triton kernels with standard PyTorch operators sized to the NeuronCore on-chip buffer and reworked the Muon optimizer’s communication, cutting per-rank traffic by 24x and raising throughput by 1.63x. On a single trn2.48xlarge instance, the model trained at about 120,000 tokens per second and served 11,544 output tokens per second across 1,280 concurrent sequences. Source