The one thing
Amazon Bedrock makes Z.ai's 753B-parameter GLM 5.3 generally available
Why it mattersA 753B open weight coding model now runs on with AWS security and compliance, so you can test it without a new vendor.
Seventy: Tuesday, 6 October 2026
What happened. Amazon Bedrock now offers Z.ai's GLM 5.3, an open-weight model aimed at coding and long software engineering tasks. It can take in about a million tokens of input and write up to 128K tokens in one reply.
OpenAI explains how it will watermark text under EU rules
Why it mattersDetection access starts with researchers, so you cannot verify yourself yet.
What happened. OpenAI published its approach to text watermarking under EU provenance rules, covering where watermarks apply and how detection works. Access to detection starts with researchers.
OpenAI adds visual ads and measurement tools for advertisers in ChatGPT
Why it mattersChatGPT now sells visual ads with attribution and brand controls, so it is a paid channel you can budget against.
What happened. OpenAI is adding a visual ad format inside ChatGPT. It is also widening the ways advertisers can track results, linking up with outside partners to credit sales to ads, and adding controls over where their ads appear.
AWS adds SageMaker inference skill for Kiro, Claude Code and Codex agents
Why it mattersYour existing coding agent can now benchmark and configure SageMaker endpoints for you, with no separate tooling.
What happened. Amazon SageMaker AI has released the aws-ai-ml skill through the Agent Toolkit for AWS. Once installed, coding agents such as Kiro, Claude Code and Codex can benchmark endpoints, recommend deployment setups, compare performance runs and write SageMaker Python SDK v3 code.
05Models
Amazon releases Nova 2.5 Sonic, a faster voice agent model, to all customers
Why it mattersAmazon's voice model now handles multi-step tasks like looking up an order and starting a return, so test it against your current voice stack.
What happened. Amazon has made Nova 2.5 Sonic generally available, a speech-to-speech AI for real-time voice agents. It follows multi-step instructions and calls connected tools more accurately, with lower response delay.
06Agents
AWS makes Anthropic's Claude 5.5 models available in GovCloud for regulated workloads
Why it mattersClaude Opus 5.5 and Sonnet 5.5 now run in AWS (US), so ITAR-bound teams can evaluate AI coding tools.
What happened. Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5 are now available in the AWS GovCloud (US) Regions through Amazon Bedrock. That opens AI-assisted coding to workloads under rules such as the International Traffic in Arms Regulations.
07Agents
AWS shows how to automate moving Amazon Quick agents from dev to production
Why it mattersMoving Amazon Quick agents from dev to production AWS accounts no longer has to be a manual, error-prone chore.
What happened. AWS published a walkthrough for automatically moving Amazon Quick assets, such as AI assistants, connectors to other systems, and their data sources, from a test AWS account into a live one. It uses an auditable, repeatable server built on Amazon Bedrock AgentCore, replacing a manual process that often went wrong.
08Agents
AWS explains how agentic retrieval fixes partial answers in LangChain and Bedrock RAG apps
Why it mattersIf your support assistant answers comparison questions, one blended search returns relevant-looking chunks that cover only part of the question.
What happened. AWS shows how to pair LangChain with Amazon Bedrock Knowledge Bases so a RAG assistant splits a multi-part question into separate searches. A single query vector blurs all the intents together, so answers cover only part of what was asked.
09Agents
AWS explains how to evaluate multi-agent AI systems with Bedrock AgentCore
What happenedAWS published a blog post on evaluating for helpfulness, accuracy and using AgentCore. The supplied text covers only the problem framing, with no figures, features or customers.
That's the whole issue.
9 stories, about 3 min 12 sec. Nothing was left out to fill the time.
Words of the day
- Context window
- How much text the model can hold in mind at once — your prompt and its reply combined. 256,000 tokens is roughly a 600-page book.
- text watermarking
- Embedding a hidden statistical pattern in AI-generated text so a detector can later identify it as machine-written.
- RAG
- Retrieval-augmented generation: look up relevant documents first, then let the model answer using them. It grounds answers in real sources instead of memory.
“We choose to go to the Moon in this decade and do the other things, not because they are easy, but because they are hard.”