Episode Details

The 90%-off model, drawn on a legal pad. Illustration: Artificially Intimidating

Good day, humans. Two things happened while you were pricing out your AI budget: a Chinese model called GLM 5.2 quietly captured a third of US enterprise traffic by charging up to 90% less, and Anthropic’s Claude Cowork slipped off the laptop and onto your phone. Also in today’s Context Window: Polymarket is betting on Anthropic to beat OpenAI to an IPO, a study caught chatbots winging your financial advice, and an MIT scholar warns we’re farming a dangerous “intelligence monoculture.” Let’s get into it.

A $0.91 Chinese Model Is Eating US AI

CNBC

What happened: Z.ai’s open-weight GLM 5.2, released June 16 under a permissive MIT license, now accounts for more than 30% of US enterprise AI traffic on OpenRouter — peaking near 46%, up from about 4.5% a year ago. It scores within a point of Anthropic’s Claude Opus 4.8 on a key coding benchmark at roughly 60–90% lower cost.

Why it matters: The single biggest cost of building with AI is the per-token bill. A frontier-grade model at a tenth of the price means cash-strapped US startups can quietly route their workloads through a Chinese lab and pocket the difference — no visa, no import, just a download.

What everyone’s saying: Vercel, Fireworks, Databricks and Microsoft all rushed to host it, while US investors and the Trump administration warn about the security of an unrestricted, Chinese-controlled model that anyone can download and run with zero oversight.

My read between the lines: The moat was never the model — it was the margin. When a model trained entirely on Huawei chips, with zero Nvidia inside, undercuts you by 90%, “responsible AI” starts to sound an awful lot like “please pay more.”

📖 Further reading: Thanks to Apple, Your Favorite AI Tool Is a Dead Tool Walking — the case that frontier models are becoming interchangeable commodities, which a 90%-off clone proves in real time.

Speaking of doing more for less: everyone’s hunting a cheaper model, but the real cost is the work itself. Viktor is an AI agent that lives in Slack and plugs into 3,000+ tools — hand it a task and it reconciles the spend report, builds the dashboard, and ships the first-draft campaign, then hands it back done. Not a chatbot you babysit — a coworker who owns the outcome. New readers get $50 off their first month. Hire Viktor →

Your AI Coworker Just Escaped Your Laptop

Anthropic

The work clocks out of your laptop and keeps going on your phone. (AI illustration: Artificially Intimidating)

What happened: Anthropic brought Claude Cowork — its agent that works autonomously across your files, email and apps — to mobile and web, so tasks keep running after you close the laptop. It also revealed that more than 90% of Cowork sessions have nothing to do with coding.

Why it matters: Until now the agent stopped the moment you walked away. Now scheduled jobs run with no device online — a 6 a.m. client briefing can build itself — and when Claude hits a decision only you can make, the question pings your phone mid-meeting.

What everyone’s saying: It’s Anthropic going straight for the non-developer knowledge worker — what it calls the “work around work” — putting Cowork head-to-head with Microsoft Copilot and Google’s workplace agents, as

Listen Now