🚨 Anthropic built a model too risky to release

This week showed what happens when AI starts colliding with money, infrastructure, work, and the physical world.

Anthropic dominated the week, but the bigger story was what this says about where AI is heading next.

Anthropic says Claude Mythos Preview is its most capable model yet, but instead of launching it publicly, the company put it behind a new defensive initiative called Project Glasswing.

According to Anthropic’s own materials, the model has already found thousands of high-severity vulnerabilities, including flaws in every major operating system and major web browser.

Anthropic says it can identify and exploit zero-days well enough that broad release would create serious cyber risk.

This is one of the clearest signs yet that frontier AI is running into capability ceilings set not by compute or demand, but by risk.

Anthropic says its annualized revenue run rate has crossed $30 billion, up from $9 billion at the end of 2025.

Reuters reported that OpenAI was at roughly $24 billion earlier this spring, which means Anthropic may now be ahead on that measure, even though OpenAI still carries the bigger brand and private valuation.

Much of Anthropic’s jump appears to be tied to Claude Code and heavy enterprise usage, where token-intensive developer workflows translate into real money fast.

For a long time, Anthropic felt like the company people respected while OpenAI felt like the company that won. That gap is starting to look smaller.

Anthropic has launched Claude for Word in beta, giving Microsoft Word users an AI assistant that can revise text while preserving layout, formatting, numbering, and style.

It can work through comment threads, add edits as native Track Changes, answer questions with clickable citations to parts of the document, and share context with Anthropic’s Excel and PowerPoint add-ins.

The product is aimed squarely at legal, finance, and other document-heavy professions.

This is where AI gets sticky: not in flashy demos, but inside the software people already spend their day living in.

Nearly half of planned US data center capacity for 2026 is now expected to be delayed or canceled, not because companies lack money or GPUs, but because they cannot get enough transformers, switchgear, batteries, and grid access.

Bloomberg reporting cited roughly 16 GW of planned capacity across 140 projects, with only about 5 GW actually under construction.

Lead times for critical electrical gear have stretched so far that some projects are stalling before they even truly begin.

This is the real economy reasserting itself.

Chinese robotics company Unitree says its H1 humanoid robot hit 10.1 m/s on a real athletics track, reclaiming the speed title for bipedal robots.

That puts it very close to the average pace of Usain Bolt’s 100-meter world record.

The company says the run relied on upgraded high-torque actuators and reinforcement learning for real-time motion and balance.

This is partly a stunt. It is also a signal.

AI is moving into a new phase where the hard questions are not just about what a model can generate.

They are about what it can be trusted to do, how it makes money, what software it quietly takes over, what hardware it depends on, and how physical it starts to feel.

That is why this week felt so consequential. Not because one demo went viral, but because the boundaries between AI, infrastructure, enterprise software, and robotics are getting thinner by the month.

If this issue helped you make sense of AI’s chaos, forward it to a friend who shouldn’t be sleeping on this.

Until next time,
Long Live AI