Mike Isaac and Cade Metz, reporting for The New York Times:

Anthropic, once the lesser-known artificial intelligence competitor to OpenAI, has been on an inexorable rise over the past few months.

The San Francisco company recently dueled with the Pentagon over the use of A.I. in warfare. It released a powerful A.I. model, Mythos, that it said was uncannily capable of finding and exploiting hidden flaws in software. And the company advised Pope Leo XIV on his papal encyclical delivered on Monday, in which he warned about safeguarding humanity from A.I.’s most disruptive effects.

On Thursday, Anthropic punctuated its ascent by officially passing OpenAI as the world’s highest flying A.I. start-up. Anthropic said it had raised $65 billion in financing that values it at $900 billion before the inclusion of the new capital, a deal that puts it ahead of OpenAI’s last valuation of $730 billion. The company also unveiled a new flagship A.I. model, Claude Opus 4.8, which is significantly better than its predecessor at generating computer code.

$900 billion on $47 billion of revenue is hysterical. The AI industry is very much in a bubble, and both Anthropic and OpenAI’s valuations will deflate significantly in the months and years after their initial public offerings, rumored for later this year. Exactly when that deflation happens is debatable, but it definitely will. Anthropic is almost a trillion-dollar company — a distinction it took Apple decades to achieve as the first to hit the milestone in 2018 — and it does not even make $100 billion in revenue. For context, Apple’s revenue run rate upon reaching a $1 trillion market capitalization was over $255 billion.

But Anthropic is unique as the only AI laboratory to turn a profit, due to its massively profitable enterprise deals and stingy rate limits. Anthropic does not like subsidizing heavy usage for its customers, as I wrote about in April. In the bubble, Anthropic might be the only properly valued company. OpenAI is struggling financially — it is unable to pay for its massive data center projects and has no financially successful product for an IPO. In a vacuum, it is OpenAI that is overvalued, not Anthropic.

So why does Anthropic “deserve” a $900 billion valuation? I’ve been thinking a lot about this as OpenAI rapidly iterates its Claude Code competitor, Codex. The OpenAI team has been shipping updates faster than any tech company I’ve seen, and every update is quite impressive. The Codex remote control feature, allowing users to code from their iPhones, is fully featured and excellent. The Computer Use plugin is astounding. The Codex plan rate limits are the best in the industry, and people can even use them with third-party tools like OpenCode and T3 Code. By comparison, Anthropic’s products are substandard. The Mac app is almost unusable, the Dispatch remote control feature is basic, and the computer use tool is nothing like OpenAI’s.

OpenAI is objectively winning the product development race. Codex is an immeasurably better product than Claude Code, and people on social media adore it. I hear almost nothing about Claude Code anymore among the developer crowd, in stark contrast to just a few months ago. Yet everyone is loath to cancel their Claude Code subscriptions — certainly enterprise customers, sans Microsoft, which is transitioning its employees to its bespoke GitHub Copilot product. While Claude doesn’t meaningfully surpass OpenAI’s models in any benchmarks — they appear to be neck and neck — the experience of using the model is absolutely better. This has created an interesting debate between models and products. OpenAI has the best products and great models, whereas Anthropic has fantastic models and terrible products.

This chasm will only widen in the coming weeks as Anthropic deploys Mythos-level models to the public. (Claude Mythos is Anthropic’s name for a much more intelligent model that surpasses the latest OpenAI and Google products.) Anthropic’s models have such an intuitive sense for what makes for great code, which is especially important to enterprise customers. While vibe-coders on social media care least about code quality, major corporations want their codebases to be workable. Fundamentally, Codex is better at working through problems and writing many lines of code to fix them, whereas Claude Code is slow, methodical, and detail-oriented. They’re just different mindsets entirely. Only one appeals to people who look at code. The Anthropic models are better at writing code — the OpenAI models are better at solving problems.

Anthropic will continue to be profitable as long as the models are more important than the products.