Jul 26, 2026
Today in brief.
Anthropic is now alone on open weights
Google and OpenAI signed the pro-open-weight letter, Jensen Huang put his name on it, and the fight stopped being industry-vs-Washington and became everyone-vs-Anthropic.
🤖 AI
Opus 5's ARC-AGI score is getting picked apart
Two days after launch the benchmark community is calling Opus 5's ARC-AGI result benchmaxxed, and independent agentic tests are painting a more mixed picture than the launch chart did.
🔧 Tech
Anthropic deleted 80% of Claude Code's system prompt
The new context-engineering rules for Claude 5 models say the same thing the Claude Code team proved in production: the elaborate prompt scaffolding you built for older models is now actively hurting you.
💼 Business
Corporate America started cutting AI budgets
The WSJ says enterprises are suddenly done overspending on AI, and two fresh economics papers say the jobs apocalypse isn't showing up in the data either - the sober quarter has arrived.
The local model stack had a very good day
llama.cpp got full MCP support, someone ran a real LLM on an $8 microcontroller, and two open-source releases attacked TTS size and KV-cache memory - the local layer is filling in fast.
🤖 AI
Apple's quiet AI position gets a second look
Apple is reportedly in talks with a model-compression startup to fit real AI on an iPhone, and the contrarian case that Apple already won on-device AI is getting traction.
No stories in this category for this edition.