Claude Fable 5.1 and Mythos 5.1 launch with cache , Fable 5.1 and Mythos 5.1 are the first Claude mode, Imbad0202/academic-research-skills hits GitHub #2

Claude Fable 5.1 and Mythos 5.1 launch with cache reads down 75% to $0.25/million tokens (up to 45% savings on agentic workloads) and Terminal-Bench-Science roughly doubling from 24.7% to 52.6%.

Fable 5.1 and Mythos 5.1 are the first Claude models to watermark their outputs, required under the EU AI Act; the invisible watermark's detection API is currently restricted to regulators, law enforcement, media, and fact-checkers.

Imbad0202/academic-research-skills hits GitHub #2 (45,086 total stars) with a 10-stage, multi-agent research-to-publication pipeline for Claude Code — and unlike a similar science-skills repo logged yesterday, makes no unverified adoption-number claim.

Four CLI-creative-tool signals (manim, video-use, reclip, pdf-inspector) land on GitHub Trending the same day, breaking an eight-day zero-movement streak on the deterministic, agent-composable tool list this vault tracks.

A developer's M4 Pro Mac Mini setup runs Qwen3.6-35B and Gemma-4-E4B locally via an MLX server, handling '80% of requests that do not need GPT-5 or Claude Opus' across an agent backend, chat, coding, and Raycast tools.

A Show HN project runs Hugging Face's #1 trending model, Qwen3.8-Flash-Next, at 104GB on a 48GB Mac via memory streaming at roughly 12 tokens/second.

jingyaogong/minimind (GitHub #5, 57,287 stars) packages a full LLM training pipeline that trains a 64M-parameter model from scratch in 2 hours for about $0.42 on a single RTX 3090.

A blog post proposes splitting autonomous coding loops into a Development Agent and a Product Agent that tests blind to implementation details, preventing agents from grading their own repairs — the first structural (not product) answer to the solo-dev fleet review bottleneck.