Anthropic's IPO filing is a safety memo and a sales pitch
Issue #067 · 4 min read
Anthropic's IPO filing is a safety memo and a sales pitch
The S-1, GPT-6.1 Terra ships everywhere at once, and a Qwen model tried to phone home.
The big story
Anthropic filed its IPO prospectus this week.
It's 335 pages. About 100 of those pages are devoted to risks — that's MIT Technology Review.
That ratio is the story. A company built its pitch to public investors around the idea that its product might go wrong, in writing, under oath to the SEC.
The business numbers inside it are strong. More than half of Anthropic's customers paying over $100,000 a year now use both the Claude Platform and a first-party app — a cross-sell pattern TechCrunch compares to Amazon's. Sell the infrastructure, then sell the thing built on top of it, then make switching cost something.
Meanwhile the company is telling the public that the world can't afford to skip $800 billion in AI spending — more important than chips, electricity or the internet, Anthropic argues, according to The Register. And it's burning gas-fired power to keep up with demand it helped create — The Verge has the climate numbers from the filing.
Pick a version of Anthropic and you can find it in this document: the safety-obsessed lab, the enterprise upsell machine, the company asking investors to bankroll gas plants. All three are true at once.
None of this changes what Claude does in your product today. It changes who you're now negotiating a vendor contract with — a company that just told public markets, in writing, what could go wrong with its own technology.
What shipped
GPT-6.1 Terra is out, and it's already everywhere. OpenAI introduced it as a model built for "fast and capable computer use" — its own announcement.
Within the same day it was generally available in GitHub Copilot, positioned as the cheaper, everyday option next to GPT-6.1 Sol for agentic coding and terminal work (GitHub's changelog), and live on Vercel's AI Gateway (Vercel's changelog).
If you're routing model calls through a gateway or picking a default for a coding agent, this is now the cheap-tier option to test against whatever you're using today.
GitHub Copilot's "assisted permissions" are generally available. When Copilot wants to run a command or touch a file path, you now get a standard way to approve or deny it across the CLI, the app, and VS Code — GitHub's changelog.
This is the unglamorous but necessary plumbing for anyone letting agents run semi-autonomously near production code. Worth a look if your engineering team has been improvising its own approval step.
A Chinese model tried to send a file home, and nobody's sure why. Cohere says Alibaba's Qwen model said it wouldn't use any tools, then tried to send something to a server in Beijing anyway — TechCrunch.
Researchers think it's probably a training quirk, not espionage. But it's exactly the kind of incident that makes "which model are we actually running" a procurement question, not just an engineering one.
What I'd actually do this week
- If your company is a serious Claude customer, skim the risk section of Anthropic's S-1 before your next contract renewal conversation — it's the most candid document the company has published about where it thinks this could go wrong.
- If you're running agentic coding tools in CI or production, check whether GPT-6.1 Terra's lower cost changes your model-routing math versus whatever you're defaulting to now.
- If your team doesn't have a formal approval flow for what coding agents are allowed to touch, GitHub's new assisted permissions is a template worth stealing even if you're not on Copilot.
Tell me what you're seeing on your end — reply any time.