The metered AI era

Hi everyone, I hope you had a nice week despite the intense heat wave that is currently happening. I’d like to thank Box once again for supporting Hypertext as its very first sponsor. Feel free to share this email with a friend, it really helps 💌 Sign up here.

The business model finally caught up with the AI model

It turns out companies can raise tens of billions of dollars before finding so-called ‘product-market fit.’ Or, at least, that’s Simon Willison’s theory.

According to him, OpenAI and Anthropic have just found product-market fit (in April 2026!). Before today, those two AI behemoths were just fiddling around to find a product that sells itself. Given the large product catalogs of the two AI labs, you might be wondering what are these miraculous products capable of generating billions in revenue.

And yes, you guessed correctly! It’s Claude Code and OpenAI’s Codex. More precisely, it’s those agentic coding products deployed with enterprise clients willing to switch to usage-based pricing.

Anthropic is about to close its first profitable quarter ever. This seems weird, given that developers can’t stop talking about Anthropic’s compute shortage. But if you run a bakery, and you run out of croissants by 10 in the morning, you have some pricing power in your hands. You can also safely increase your croissant batches.

And that’s precisely what Anthropic has been doing over the past month:

1. Enterprise customers renewing their contracts now pay $20/seat/month plus tokens instead of a more traditional SaaS-like fixed monthly price. OpenAI moved Codex to token-based pricing in April, too. Some companies are now paying hundreds, sometimes $1,000+, per software engineer on Claude Code or Codex.
2. Anthropic has been investing as much as possible in compute capacity.
3. The more compute it can secure, the more revenue it generates.

So I’m sure Anthropic won’t stay profitable for long… They’ll most likely reinvest all that money in more infrastructure, given how the company’s revenue chart basically looks like a pro-level climbing route (i.e., vertical).

SPONSORED by Box

Huge amounts of business content remain ungoverned. Box brings structure, security, and control to your unstructured data, enabling secure, enterprise-ready AI.

Learn more

The agent is eating the chatbot

What about you, ChatGPT? OpenAI is the clear leader in consumer AI with 900 million weekly active users. My back-of-the-envelope calculation still puts it above Codex in terms of revenue. But Codex is catching up pretty quickly.

But more important than pure numbers, the vibe is changing at the company. Based on conversations I’ve had, Codex used to be more or less a startup within a startup, with a separate team trying to ship new features as fast as possible without having to worry too much about ChatGPT’s massive installed user base.

And it’s been working so well that the Codex team is gaining internal political power within OpenAI. Thibault Sottiaux, the former head of Codex, now leads product and platform across:

• Consumer (ChatGPT).
• Enterprise (historically ChatGPT Enterprise, now increasingly Codex).
• Developer products (historically APIs, now increasingly Codex + APIs).

The product teams behind ChatGPT, Codex and APIs are also being merged. My reading is that the Codex team is slowly taking over OpenAI’s entire product portfolio. In a not-so-distant future, people might end up saying “OpenAI, the company behind Codex” instead of “OpenAI, the company behind ChatGPT.”

Mistral is reaching the same conclusion from the opposite direction, as it already focuses heavily on B2B enterprise clients. This week, Mistral held its first conference here in Paris. Other than some big new clients (Airbus, EDF and BMW) and a new data center in Les Ulis (next to where I grew up), Mistral announced a product shift that is right in line with OpenAI’s moves.

Until this week, Mistral had a consumer/prosumer product called Le Chat and a developer-centric product called Vibe with a command-line interface (CLI). Mistral is merging all of those products into a single brand called Vibe. Le Chat, as a brand, is gone.

In addition to this naming adjustment, Mistral is adding support for skills and automations in a separate “Work Mode.” So now you have Vibe (the chatbot), Vibe’s Work Mode (a bit like Claude Cowork) and the Vibe CLI for developers (and a new VS Code plugin).

This new product positioning should work well for signing new customers who have yet to use AI at scale. But will that be enough to convince large enterprise customers to switch from Claude to Vibe?

Europe redraws the marketplace line

In non-AI news, Temu just became a test case for Europe’s Digital Services Act (DSA) as the European Commission fined the Chinese shopping app €200 million for failing to properly assess and reduce the risk of illegal products being sold on its marketplace.

A few months ago, Shein made a deal with the iconic BHV department store here in Paris. Reporters quickly pointed out that you can find extremely dubious items like sex dolls with childlike features, machetes and other weapons. So I’m expecting more fines pretty soon.

More importantly, the bigger point is around the DSA. The law doesn’t make marketplaces automatically liable for every illegal product sold by a third-party seller. But Brussels is clearly testing a tougher idea: if your marketplace design, recommendation engine and seller controls make illegal products easy to distribute at scale, you can’t hide behind the idea that you’re just a neutral shelf.

In practice, that starts to look a lot like responsibility for what gets sold there.

Have a good day ☀️
Romain