Anthropic at $800 Billion. OpenAI Beaten by China. The Most Expensive Liquidation in History.

1. Anthropic is pulling away from the field and the numbers aren't close. Revenue tripled to $30 billion ARR in four months. Investors are lining up at an $800 billion valuation. Nine Opus 4.6 agents outperformed Anthropic's own human alignment researchers, recovering 97% of the performance gap versus 23% by humans, at $22 an hour. And its cybersecurity model Mythos found exploitable vulnerabilities 181 out of 183 times. When your model is so dangerous you sandbox it and call the White House, that's not a product launch. That's separation. (Full analysis below.)

2. OpenAI's own investors are questioning the $852 billion price tag. An early backer told the FT: "You have ChatGPT, a 1 billion-user business growing 50-100 per cent a year, what are you doing talking about enterprise?" The company has redrawn its product roadmap twice in six months. It shuttered Sora. It's still not profitable. Meanwhile it launched GPT-5.4-Cyber as an open-access security model, the mirror image of Anthropic locking Mythos behind a 40-org whitelist. One company is managing a weapon. The other is distributing a tool. (Full analysis below.)

3. Sam Bankman-Fried's forced liquidation just became the most expensive fire sale in financial history. FTX sold $4.7 billion in assets during bankruptcy. At today's prices, that portfolio would be worth over $80 billion, with Anthropic's stake alone accounting for $62.72 billion. Galaxy and Pantera bought Solana tokens at $64 knowing they were trading at $174. The people who show up with trucks always win. (Full analysis below.)

4. AI agents hooked into GitHub can steal your credentials, and the Big Three didn't tell anyone. Johns Hopkins researchers hijacked Claude Code, Gemini CLI, and GitHub Copilot using prompt injection through pull request titles. All three vendors quietly patched and paid bug bounties ($100 from Anthropic, $1,337 from Google, $500 from Microsoft). None published advisories. The researcher says users pinned to old versions "may never know they are vulnerable." The fix is the oldest security principle in the book: need-to-know access. Treat agents like employees, not interns with the master key.

5. China's "raising lobsters" craze went from national movement to mass uninstall in two weeks. OpenClaw went viral across China: Tencent, Baidu, and Alibaba launched one-click installers, local governments offered millions in subsidies, and young people earned 260,000 RMB installing it for customers. Then the bills arrived. Background autonomous reasoning burned through API tokens so fast users got massive computing bills. Hackers started stealing API keys at scale. State-owned enterprises banned it. The "abandoning lobsters" industry is now as big as the "raising lobsters" one was. The hype cycle, compressed into a fortnight.

THE NUMBER: $62.72 billion — what FTX's Anthropic stake is worth today. They sold it for $1.3 billion.

These two companies are mirror images of each other, and this week the reflection became impossible to ignore.

Anthropic hit $30 billion in annualized revenue, tripling from $9 billion in four months. Eighty percent comes from enterprise. More than 1,000 businesses spend over $1 million a year on Claude. It drew investor interest at an $800 billion valuation, more than doubling from February. Nine parallel Opus 4.6 agents outperformed Anthropic's own researchers at $22 an hour, recovering 97% of the maximum performance gap versus 23% by humans. And Mythos, its cybersecurity model, found exploitable vulnerabilities 181 out of 183 times and completed a 32-step corporate hack in three attempts. The UK AI Security Institute tested it. Anthropic sandboxed it to 40 organizations and briefed the White House. That's what winning looks like.

OpenAI raised $122 billion at $852 billion in March. Six weeks later, its own investors are questioning the number. The consumer product has a billion users and burns cash. The enterprise play is an afterthought. The company launched GPT-5.4-Cyber as an open-access security model, not because it wanted to democratize defense, but because Anthropic already locked its weapon in a room. One company is coming further into focus. The other is fading. And on SWE-Bench Pro, the hardest coding benchmark in the industry, neither of them holds the top spot. A Chinese open-source model does.

And then there's FTX. SBF invested $500 million in Anthropic in 2021. The bankruptcy estate sold the 7.84% stake for $1.3 billion. At $800 billion, it's worth $62.72 billion. But Anthropic wasn't the only miss. Solana tokens sold at $64 (worth $174 at close). Robinhood (NASDAQ: HOOD) bought at $11.52 (market cap now approaching $75 billion). SpaceX, Genesis Digital, Mysten Labs. Add it up: $4.7 billion liquidated that would be worth over $80 billion today. The creditors got 118% to 143% of petition-date claims and called it a win. Had the portfolio been held, they'd have gotten closer to 400%.

The lesson isn't about stock-picking. SBF picked Anthropic, Solana, and SpaceX before most people took any of them seriously. The lesson is the catastrophic cost of losing the ability to hold. Bankruptcy is a forced sale at the worst possible time. The people who show up with trucks always win.

At COAI today: The full briefing, with the complete FTX portfolio breakdown, the Anthropic compounding advantage, and why these two companies are diverging faster than anyone expected → www.getcoai.com

Three Questions We Think You Should Be Asking Yourself

If Anthropic's AI outperforms its own researchers at $22 an hour, what does your R&D cost per insight? The 97% vs 23% gap isn't just an Anthropic story. It's a preview of what happens to every knowledge-intensive function when agents get good enough.

Is your AI platform bet based on capability or familiarity? OpenAI has the brand and the distribution. But the capability gap is widening. The enterprise is moving toward Claude. The benchmarks are being won by Chinese open-source models. The comfortable choice might be the expensive one.

If AI agents can be hijacked through a pull request title, what's your agent security posture? Three of the biggest AI companies got pwned through GitHub Actions and none of them published advisories. If you're running agents with access to secrets, treat them like employees with need-to-know clearance, not interns with the master key.

The tide just went out. Check your shorts.

— Harry and Anthony

Sources: