Anthropic filed confidentially for a public offering this week at a $965 billion post-money valuation — making it the most valuable private AI company in the world and nudging past OpenAI’s $852 billion mark for the first time. The number is attention-grabbing and slightly vertiginous: Anthropic is not profitable, frontier model training costs billions, and the half-life of any competitive advantage is measured in months. What supports the premium is mostly two words: Claude Code.
The coding tool has become the fastest-growing product in Anthropic’s history. The company’s annualized run rate crossed $47 billion earlier this month, fueled by enterprise engineering teams who have made Claude Code their default environment. Anthropic expects its first operating profit this year and is targeting an October NASDAQ listing that would rank among the largest technology debuts in Wall Street history. The $65 billion Series H round that preceded the filing — co-led by Altimeter, Sequoia, and others — was itself the largest single private funding round ever completed by an AI company.
The irony is that the engineers driving Anthropic’s valuation are also the ones getting their first taste of what AI tooling actually costs. On June 1, GitHub switched Copilot from flat-fee subscriptions to usage-based billing in “AI Credits” — one credit per cent, consumption measured in tokens. Within hours, users reported burning through weeks of credits in a single agentic session, with some Pro+ subscribers projecting their monthly quota depleted in under two days. GitHub’s position is defensible: Copilot does substantially more than it did two years ago, and metered billing honestly reflects the compute it consumes. But the psychological transition from unlimited flat-fee to a ticking meter is brutal, and it landed at an uncomfortable moment.
The uncomfortable moment being that Uber had just disclosed it burned through its entire 2026 AI tooling budget in four months. Some engineers were running individual monthly bills of $500 to $2,000 in token consumption. The company has now capped all employees at $1,500 per AI coding tool per month, and Uber’s COO has openly questioned whether the productivity returns justify the cost. That question — so far largely avoided in public — is one enterprises are starting to ask seriously.
Microsoft’s announcements at Build may be a partial answer. The company unveiled MAI-Code-1-Flash, a 5-billion-parameter coding model integrated directly into GitHub Copilot and VS Code, alongside MAI-Thinking-1, a reasoning companion. The stated goals were efficiency and reduced dependency on external models. The unstated goal is probably cost: if Copilot now charges by the token and each token is purchased from someone else, the cheapest token is the one you mint yourself. Microsoft is building the capacity to reduce its OpenAI bill. This matters because it signals that the model-as-margin dynamic driving AI lab valuations is under quiet pressure from the biggest distribution channel in the ecosystem.
The policy backdrop is gentler than it might have been. Tuesday’s White House executive order asks frontier labs to voluntarily submit their most powerful models for a 30-day government review before public release. An earlier draft had proposed a 90-day mandatory review; industry objections and White House concerns about competitiveness whittled it down to something labs can simply decline to participate in. The order does mandate AI cybersecurity benchmarks and a government clearinghouse for AI vulnerability information, which are non-trivial. But on the core question of whether the government gets a meaningful look at frontier models before deployment, the answer for now is: only if labs choose to cooperate.
The week’s underlying tension: extraordinary value is concentrating at the top of the AI stack while real costs are dispersing all the way down to individual developer sessions. The economics are going to keep moving until they find an equilibrium. Nobody knows what that looks like yet.