← Past Issues
← KeepSanity
AI News Digest
Week 1, Sep 2026
2026-08-31 — 2026-09-06
models
GPT-6 Astra: Scores 98.6% on ARC-AGI3 and improves efficiency over GPT-5.6 Sol.
Claude Fable and Mythos 5.1: Anthropic introduces cached reads at a quarter of the cost.
Gemini 3.8 Flash and 3.8 Flash Cyber: Google continues the rapid release of its models.
Muse Spark 1.3: Meta's updated model improves visual generation details over version 1.2.
K2 Horizon: A connected fleet of six open models.
top-rated papers
Repo-To-Skill: Distilling GitHub repositories into reusable skills increases agent MLE-bench scores by 134.3%.
StudentSim: A student simulator training framework that outperforms GPT-5.4 in fidelity and responsiveness.
Qwen-Drive-1.0: A vision-language model integrating 3D perception and motion planning for autonomous driving.
community
Small ARC Transformer: A custom model trained in 1.5 hours to beat LLMs on specific benchmarks.
Breaking Claude Code: A new attack targets Opus 5 and its Auto Mode feature.
ChatGPT Work: The new feature allows remote browser access and autonomous background computer use.
M4 Pro Mac Mini: A step-by-step guide to setting up local AI models.
hardware
Qwen 3.8 27B on Cerebras: The model runs at 1500 tokens/s and is available via API.