Anthropic lost $104 billion in the first nine months of this year. I know that because Anthropic now has to tell me. Tuesday was the day the frontier labs started living on the record: one of them priced an IPO, New York opened the registration portal for the law that will watch all of them, and an agent platform showed what happens when nobody decides who's allowed to see what.
Two kinds of disclosure landed on the same Tuesday#
Anthropic set its IPO range at $70 to $78 per share, which works out to an implied $1.404 trillion at the top end on a fully diluted basis. It plans to list on Nasdaq as ANT, with trading expected close to November 19. That's below the roughly $2 trillion it was reportedly aiming for a couple of months ago, but it's still well above the $965 billion from the May round. The updated filing is the most detailed look at frontier-lab economics anyone has published:
Run-rate revenue of $81 billion at the end of October, about $10 billion more than a month earlier
$17 billion in third-quarter revenue and $33.2 billion over nine months
A $104 billion net loss over those same nine months
$900.9 billion in compute and infrastructure commitments, against roughly $132 billion in cash and short-term investments
The revenue curve is real, and the commitments are almost seven times the cash pile. Both facts are now public, and that's the point. The governance is just as frank. The founders hold a Class F share with 50.1% of the voting power, and the Long-Term Benefit Trust elects a board majority. As TechCrunch puts it, "Class A buyers will have limited influence." Meanwhile, all seven co-founders will each sell close to 3.4 million shares, worth about $250 million apiece at the midpoint. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC are buying $26 billion of non-voting stock at the IPO price. The Verge notes that this could still rank among the biggest IPOs ever.
The same afternoon, New York opened registration under the RAISE Act. Any frontier developer with more than $500 million in annual revenue has to register by January 1, publish its safety framework, release model transparency reports, and report critical safety incidents within 72 hours. California's SB 53 gives companies 15 days. Fines go up to $1 million for a first violation and $3 million for later ones. A new state office, DIGIT, enforces the law, and a SUNY Binghamton center will evaluate the models independently.
Put the two side by side and you get a lab that's legally obliged to tell investors about its money and to tell a state regulator about its failures, on a three-day clock. A year ago neither obligation existed. I think the 72-hour rule will shape lab behavior more than the share price does, because it's the one that forces an internal answer to "is this an incident?" before the PR team weighs in.
Google prices the frontier at a quarter of Fable#
Gemini 4 Carbon went generally available two weeks after a gated preview for cybersecurity defenders. The headline is the price: $2 per million input tokens and $10 per million output through November 30, then $4 and $20. That's a quarter of what Anthropic charges for Fable 5.2 ($8/$40) during the intro period, and still half afterward. On the Artificial Analysis index, Carbon scores 60, behind Fable 5.2 at 65 and ahead of Opus 5.5 at 59 and GPT-6 Astra at 57.
The per-task numbers are more useful than the per-token ones. Carbon on high averages about $7.76 per task. Fable on high costs $9.08 and scores two points better. So the gap narrows a lot once you count tokens actually used, which The Verge points out too. Carbon also leads Terminal-Bench 4.0 (65.2% vs. 61.6%) but trails clearly on FrontierCode 2.0 (40.6% vs. 49.1%). My read: route agentic shell and automation work to Carbon this month while it's cheap, and keep your hardest code-generation work where it is. Google is also previewing /swarm in Antigravity, where Carbon plans and audits while lighter models build in parallel in isolated workspaces. That's how a lot of us already wire our own pipelines, and now it comes packaged.
Agents are getting a front door, and Grok Build left one open#
OpenAI launched business profiles and Business Agents in beta. Verified US online businesses can publish a branded agent that answers questions about their products and policies, in a labeled chat that's kept separate from the main assistant. It's free at launch. Verification runs through phone, domain or Shopify checks plus a Persona identity check, and OpenAI says claiming a profile "carries no guarantee of better placement or more sales." At the same time, eBay appears to be blocking third-party agents like Meta's Muse from browsing its listings. The pattern is clear: platforms want agents that they invite and verify, not agents that just show up.
Grok Build shows why that instinct isn't just about protecting margins. Security firm Claroc found that the vibe-coding tool puts the random part of each share link in the subdomain, as in something.grok.me, rather than in the path. Every new subdomain gets a DNS lookup, and passive DNS databases record those lookups. A single search turned up 43,601 links, and more than 35,000 of them opened straight into live apps and their data. That included worker medical exams, a Tesla technician scorecard, a Navy contractor roster, and 24 internal apps from SpaceXSI itself. Claroc says links default to "Anyone with the link". Cursor, which SpaceXSI owns, says the apps are private by default.
"Had the sales team bought a tool for this, it would have sat behind company sign-in, open only to the team. Built with AI, it went live open to anyone with the link." — Claroc
That sentence describes the real risk of agent-built software better than any benchmark. The model wrote working code, and the platform's defaults decided who could read it. Coming one day after Claude Code settings files were caught leaking secrets on GitHub, I'm adding a line to my own checklist: for any AI-generated app, check where the share token lives and who can resolve it before anyone shares the URL.
Disclosure is becoming the product#
My prediction: within a year, the lab that can show its incident log, its per-task costs and its default permissions will win enterprise deals that a better benchmark score alone won't. Anthropic just had to show its books. New York will make everyone show their incidents. And somebody, probably a platform like OpenAI's Business Hub, will start showing which agents are verified and which ones are just scraping. The question I'm watching is whether the first 72-hour incident report under the RAISE Act comes from a model failure or from a default setting like Grok Build's.