Skip to content
Sign in
Artwork for AI This Week with PlsAutomate
AI This Week with PlsAutomate · Thursday · 21 min

Cheaper Models, Safer Sandboxes

AI agents got cheaper and better contained on the same day. Here's the AI news from October 7, 2026: Anthropic released Claude Haiku 5.5: about 75% cheaper than Haiku 4.5 on average, with adjustable reasoning for agent subtasks. Microsoft pushed local AI with Windows Hybrid Intelligence and a Surface laptop that can run 120B models locally. GitHub made local sandboxing for Copilot agents generally available, so the operating system blocks what the agent isn't allowed to touch. Liquid AI released open decision models that make choices in as little as 8 ms, with no text generation. NVIDIA explained how it trained Nemotron to gold level at the math and programming olympiads. Meta, Google and the U.S. government expanded a $1.8B AI biology initiative. Nous Research raised $90M at a $1.5B valuation for its open-source Hermes agent. GitHub added local Ollama models to Copilot CLI and a new model for catching leaked secrets. Lawsuits test whether ChatGPT counts as a product or a service, and the IMF warns about AI investment risk. The takeaway: Telling an agent "don't touch that" is not security. Real limits are enforced by the system around the agent, and every action gets logged.

0:00-21:59
Play in

transcript

No transcript — this publisher did not publish one.

show notes

AI agents got cheaper and better contained on the same day. Here's the AI news from October 7, 2026:

  • Anthropic released Claude Haiku 5.5: about 75% cheaper than Haiku 4.5 on average, with adjustable reasoning for agent subtasks.
  • Microsoft pushed local AI with Windows Hybrid Intelligence and a Surface laptop that can run 120B models locally.
  • GitHub made local sandboxing for Copilot agents generally available, so the operating system blocks what the agent isn't allowed to touch.
  • Liquid AI released open decision models that make choices in as little as 8 ms, with no text generation.
  • NVIDIA explained how it trained Nemotron to gold level at the math and programming olympiads.
  • Meta, Google and the U.S. government expanded a $1.8B AI biology initiative.
  • Nous Research raised $90M at a $1.5B valuation for its open-source Hermes agent.
  • GitHub added local Ollama models to Copilot CLI and a new model for catching leaked secrets.
  • Lawsuits test whether ChatGPT counts as a product or a service, and the IMF warns about AI investment risk.

The takeaway: Telling an agent "don't touch that" is not security. Real limits are enforced by the system around the agent, and every action gets logged.

more episodes

All episodes