This was the week the masks came off. Not gradually - all at once. Anthropic got blacklisted by the Pentagon and gained a million users a day. OpenAI shipped its most powerful model while employees protested its military contracts. Block fired 40% of its workforce and the stock went up 17%.
If you're still thinking about AI as a productivity tool, you're reading the wrong story. This week proved it's a force that's reshaping power, labor, and ethics simultaneously. Here's what actually matters.
---
The Anthropic Paradox: Blacklisted by Government, Rewarded by Market
The timeline reads like fiction. Anthropic signed a Pentagon contract in 2025 with two red lines: no autonomous weapons, no mass surveillance of US citizens. Defense Secretary Hegseth called those conditions "woke." The Pentagon issued a 5pm ultimatum. Anthropic said no.
What happened next was swift. Trump ordered federal agencies to phase out Claude within six months. Hegseth officially declared Anthropic a "supply chain risk." Defense contractors started dropping Claude from their stacks.
The same week, Anthropic CPO Mike Krieger (yes, the Instagram co-founder) revealed Claude is getting over 1 million new signups per day. The #QuitGPT movement sent waves of ChatGPT users to Claude. ZDNET and Inc published migration guides. Claude even added a memory import feature to make switching easier.
Meanwhile, Dario Amodei's leaked memo called OpenAI's Pentagon deal "safety theater" and named Palantir and Sam Altman directly. The Under Secretary of War tweeted back calling Amodei a "liar" with a "God complex." OpenAI quietly revised its contract to add the same surveillance guardrails Anthropic demanded from the start.
Read that again: Anthropic got blacklisted for asking. OpenAI got praised for eventually agreeing. Same guardrails, different PR.
The business lesson is counterintuitive but clear. Anthropic's principled stance didn't just protect its brand - it became a growth engine. Ethics as a moat. The government punishes you for saying no to weapons. The market rewards you for the exact same reason.
1 million signups per day. While blacklisted. That's not just growth - that's a mandate from the market about what it values.
Takeaway: In AI, trust is becoming the ultimate product differentiator. You can't A/B test your way to integrity. Anthropic is proving that doing the right thing and building a great business aren't competing priorities - they're the same priority.
GPT-5.4: AI Just Passed Humans at Using a Computer
OpenAI dropped GPT-5.4 on Wednesday - and the headline number is genuinely startling. On the OSWorld-Verified benchmark (real desktop navigation using screenshots and keyboard/mouse control), GPT-5.4 scored 75.0%. The human baseline? 72.4%.
Let that sink in. AI is now better than the average human at literally clicking around a desktop. Not at thinking. Not at creating. At navigating a computer.
The jump is what makes it scary: GPT-5.2 scored 47.3%. One generation later, it's 75%. That's not incremental improvement - that's a capability wall being demolished.
The ecosystem response was equally telling. Cursor integrated GPT-5.4 within hours, calling it "leading on internal benchmarks." Perplexity, Windsurf, Notion - same-day adoption across the board. A model launch that would have taken months to propagate now takes hours. That's how mature this ecosystem has become.
But here's the nuance most coverage missed: Cursor also noted that Claude Opus 4.6 still leads on "complex multi-file architecture and agentic teams." GPT-5.4 wins on raw benchmarks, financial plugins, and 1M token context. Claude wins on deep coding and long-form reasoning.
We're entering an era where no single model wins everything. The real competitive advantage isn't picking the right model - it's knowing which model to reach for which task. Model loyalty is dead. Workflow wins.
Takeaway: The benchmark numbers are impressive, but the real story is the speed of ecosystem adoption. Same-day multi-platform integration is the new standard. If your product can't absorb a new model in hours, you're already behind.
The Jobs Story Nobody Wants to Hear
Three data points converged this week to paint a picture most people aren't ready for.
First: Block (Jack Dorsey's company behind Square and Cash App) cut 4,000 employees - roughly 40% of its workforce. The reason, stated plainly in Dorsey's letter to shareholders: "intelligence tools." Not restructuring. Not market conditions. AI. The stock jumped 17%. Wall Street's message was brutal and clear: smaller teams, bigger margins.
Second: Anthropic released its AI Exposure Index, quantifying which professions are most vulnerable to LLM automation. Programmers topped the list at 75% task exposure. The company building the tool is telling its biggest customers they're the most replaceable. That's either radical transparency or the most savage marketing move in history. Probably both.
Third: AI progress trackers now predict over 100-hour autonomous work horizons by year's end. Current estimate: about 43 hours. The doubling rate is every 3.8 months. By December, you could hand an AI a task and come back four days later to a completed result.
Connect these dots: companies are cutting 40% of headcount because of AI. The AI makers are saying 75% of programming tasks are automatable. And the autonomous work window is doubling every quarter.
Meanwhile, someone in Japan automated bookkeeping for 60 companies per day using Claude. One accountant doing the work of a 10-person team. Not a demo - running in production right now. We keep debating whether AI will replace developers while it's quietly eating accounting, legal review, and every back-office function that involves pattern recognition and rule following.
The boring jobs always go first. And nobody notices until the layoffs hit.
Takeaway: The question isn't "will AI take my job." It's "am I the human reviewing AI's output, or the human being reviewed?" The gap between AI-native workers and everyone else is widening every week. Adapt now or adapt later - but later gets more expensive every quarter.
Quick Takes
New York's S7263 bill wants to ban AI chatbots from giving advice in 14 licensed professions including medicine, law, and engineering. This landed the same month OpenAI, Anthropic, and AWS all launched healthcare AI products. Three of the biggest companies in tech are betting billions on healthcare AI. One state is trying to outlaw it.
Cursor crossed $2B ARR - doubled from $1B in just three months. Vibe coding isn't hype anymore; it's a $2B/year industry with enterprise customers lining up.
Netflix acquired Ben Affleck's AI startup InterPositive for post-production automation. Not replacing actors or generating synthetic performances - automating the boring parts of filmmaking like color, relighting, and VFX. Hollywood is learning that AI works best when it handles the tedious stuff, not the creative stuff.
Someone vibe-coded an ad generator that turns one competitor ad + a product photo + brand kit into 100+ Facebook ad variations. A marketing team's weekly creative sprint just became a 20-minute prompt session.
This was the week AI stopped being a tool story and became a power story. Government vs. market. Ethics vs. contracts. Smaller teams vs. bigger output.
The builders who win from here aren't the ones with the best prompts. They're the ones who understand which side of every one of these shifts they want to be on.
See you next week.