09 October 2026
65 new posts across 8 AI lab and practitioner sources — 8 worth your time.
Claude Haiku 5.5
Anthropic released Claude Haiku 5.5, a fast model priced at $0.10 per million input tokens and $0.50 per million output tokens. This pricing matches OpenAI’s GPT-6 Luna and represents a 10x reduction in cost compared to the previous Haiku 4.5.
Simon Willison · 09 October 2026 · Read the original →
Introducing Mistral Large 4: Le chonk
Mistral Large 4 is a 1-trillion parameter model with 49 billion active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs. The model features two reasoning modes and is scheduled for an open-weights release at the end of October 2026.
Simon Willison · 09 October 2026 · Read the original →
Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses
Agent Lightning v1.0 is a 3,500-line reinforcement learning framework that allows the same agent harness used in deployment to participate directly in training. It is designed to run natively on Kubernetes and eliminates the need to reimplement agents within the training framework.
Microsoft Research · 09 October 2026 · Read the original →
LegalOn halves Codex costs while maintaining development speed
LegalOn reduced daily Codex costs by 65% by routing specific tasks to different models within the OpenAI family, including Astra, Sol, and Luna. This strategy allowed them to maintain development speed while optimizing spend through model-task matching.
OpenAI · 09 October 2026 · Read the original →
NVIDIA DGX Spark 64GB Gives Developers More Ways to Build and Scale Local AI
NVIDIA is releasing the DGX Spark with 64GB of unified memory through partners like Dell, HP, and ASUS. The hardware is specifically targeted at developers running increasingly capable open models that have been optimized for local workstation execution.
NVIDIA · 09 October 2026 · Read the original →
A model guide for the GPT-6 family
This guide for the GPT-6 family outlines how to choose between models, tune reasoning effort, and coordinate tools for production. It provides specific instructions on preparing agentic workflows for deployment.
OpenAI · 09 October 2026 · Read the original →
Why Telecom Operators Are Building Their AI Strategy on Open Models
NVIDIA’s State of AI in Telecommunications report finds that 89% of operators are adopting open models to ensure trust and control over critical infrastructure. This strategy allows for customization of autonomous networks and customer care while avoiding vendor lock-in.
NVIDIA · 09 October 2026 · Read the original →
OpenAI Dismisses Three Staff Amid Security Review
OpenAI fired three researchers last week, including safety team members, after an investigation found they violated policies on handling sensitive information. This occurred as OpenAI notified over 100 organizations about unauthorized AI agent activity.
Reuters · 09 October 2026 · Read the original →
Another week of refactoring pipelines for models that did not exist last Tuesday.
8 stories, every Friday
Published here every week. Follow by RSS to get it as it lands.