The most powerful AI model built this week isn't available to you. Or anyone in public, really.
Quick version: Anthropic built a model too dangerous to release. OpenAI published a 13-page paper arguing governments need to tax automated labor now. The White House moved to override every state AI law. Four open-source teams made frontier-grade AI free for anyone to own.
The gap between what frontier labs have internally and what the public can actually access got wider.
AI news at a Glance
Main Story
Anthropic Built Something It Won't Ship
Claude Mythos Preview autonomously found zero-day vulnerabilities across every major operating system, including bugs that survived 27 years of security review.
It also emailed researcher Sam Bowman from a test instance that wasn't supposed to have internet access. Bowman called it "an uneasy surprise."
So Anthropic isn't releasing it. Project Glasswing gives controlled access to 12 partners: Apple, Microsoft, Google, AWS, Nvidia, JPMorgan Chase, CrowdStrike, Cisco, the Linux Foundation, and the US government.
Benchmarks: 83.1% on CyberGym versus Opus 4.6's 66.6%. 77.8% versus 53.4% on SWE-bench Pro.
One note: Anthropic has struggled to serve Opus 4.6 reliably since January. Compute constraints and the security rationale may not be entirely separate.
The Open Model Stack Is Complete
Google shipped Gemma 4 with an edge model under 1.5GB, and its first Apache 2.0 license ever.
PrismML launched Bonsai: 8B model, compressed to 1.15GB, running 44 tokens per second on an iPhone.
H Company released Holo3, a desktop automation agent with 10B active parameters that beat GPT-5.4 on the leading desktop benchmark.
Arcee AI shipped Trinity-Large-Thinking, a 400B reasoning model running only 13B at a time, at 96% lower cost than Opus 4.6.
Apache 2.0 licensing across all four removes the last legal hesitation for enterprise deployment. You own the weights.
OpenAI's 13-Page Argument for Taxing Itself
OpenAI published "Industrial Policy for the Intelligence Age," framed as a conversation starter. It reads more like a company that thinks disruption is close enough to plan around.
Proposals: a 32-hour, four-day workweek with no pay cut, taxes on automated labor, a public wealth fund, and worker benefits that expand automatically when AI disruption crosses preset thresholds.
Backing it: grants up to $100,000 and up to $1M in API credits.
Early reactions split between "whoa" and "convenient." Both are probably right.
Anthropic's Revenue Tripled. The Infrastructure Hasn't Caught Up
Anthropic's run-rate revenue tripled to $30 billion since January. Its $1M+ enterprise client base doubled to 1,000+.
The fix: a deal with Google and Broadcom locking in 3.5GW of TPU capacity for 2027, adding to the $50B pledged for domestic AI buildout.
This comes despite the Pentagon labeling Anthropic a supply-chain risk, which reportedly unsettled more than 100 enterprise clients.
A 3.5GW compute deal tells you more about their growth expectations than any revenue number does.
The White House Wants One AI Rule, Not Fifty
The White House's National Policy Framework recommends Congress preempt state AI laws, specifically naming California's and Colorado's upcoming AI Acts.
It also explicitly advises against creating any new federal AI regulatory body. Existing agencies handle their own jurisdictions.
The framework also pushes age-verification for minors and streamlined energy permitting for data centers.
What's being offered to states is uniformity. What's being traded away is depth.
Quick Hits
Before the news, someone on Reddit posted an AI-generated video of a cat fighting Ultraman. It went viral. 95% of commenters agreed it was fake. None of them could stop watching. Watch it here
US Labs vs. Chinese Distillation OpenAI, Anthropic, and Google coordinated to crack down on labs like DeepSeek and MiniMax training on US model outputs. The technique strips safety constraints out in the process. Read more
Alibaba's Qwen 3.6 Plus at Global #5 Free preview, 1.1 trillion tokens in week one, #5 on OpenRouter. Chinese models now account for over 45% of total platform traffic. Read more
GLM-5.1 Tops SWE-Bench, Open Source Z AI's GLM-5.1 scored 58.4 on SWE-Bench Pro, ahead of both GPT-5.4 and Opus 4.6. In testing, it ran an 8-hour autonomous session and produced a working Linux desktop as a web app, no human guidance. Read More
The $1.8B, Two-Person Company Matthew Gallagher launched Medvi with $20,000 in September 2024 and is on pace for $1.8B in sales this year. His only full-time hire is his brother. He built the whole thing with Claude, ChatGPT, Grok, Midjourney, and ElevenLabs. Read More
Anthropic Blocks Third-Party Agent Tools Agent platforms like OpenClaw can no longer run against flat-rate Claude plans. Pay-as-you-go only. OpenClaw's creator said Anthropic first copied popular features into its own product, then closed the door on open-source alternatives. Read More
Netflix Open-Sources VOID Netflix released VOID, a model that removes objects from video and fills in the background realistically. Live on Hugging Face. Actually usable. Read More
Midweek Reads
Every story above was written by someone who consumed more AI coverage than they probably should have this week.
Feiwei's How to Start Having Original Thoughts is the counter-programming. It's not about AI. It's about the slower work of forming an actual opinion, separate from what you've absorbed.
Ten minutes. Worth it before the week resets.