What Simon Willison Means When He Says OpenAI and Anthropic Finally Found PMF

ChatGPT crossed 100 million users in two months. Claude has tens of millions. So when veteran developer and longtime LLM-watcher Simon Willison said recently that OpenAI and Anthropic have finally found product-market fit, the obvious response is: wait, didn’t that happen years ago? His point is sharper than it sounds. The PMF he’s talking about isn’t about users. It’s about a business that actually works.

Popularity isn’t PMF

PMF is usually a startup-stage idea. Did you build something a market wants enough to sustain a business? On the surface, ChatGPT settled that question in late 2022.

But Willison’s framing reframes the goalposts. Hundreds of millions of free users on a consumer chatbot isn’t PMF in any meaningful financial sense — it’s a giant unpaid demo running on the most expensive infrastructure in tech history. The real question is whether these labs have found a product that pays for the GPUs, the training runs, the data centers, and still leaves something on the table. For most of the last three years, the honest answer was no.

What changed is that an answer finally showed up. And it wasn’t the chatbot.

The answer was coding tools

The product that actually closes the loop is developer tooling. Claude Code, Cursor, GitHub Copilot, OpenAI’s Codex — this is where the labs are printing real revenue, and there are clear structural reasons why.

Developers will pay. $20, $100, even $200 a month is a rounding error against a software engineer’s salary. The value is legible — code either runs or it doesn’t, so you find out immediately whether the AI helped. And the labor it offsets is expensive. A senior engineer billed at $150 an hour who saves thirty minutes a day generates roughly $15,000 a year in time value. A $200/month subscription is a trivially good trade.

Anthropic is the cleanest example. Claude Code took over developer Twitter and Hacker News almost overnight after launch, and a meaningful chunk of Anthropic’s revenue reportedly flows from coding and agentic-coding use cases. The company that branded itself around safety is, commercially, a developer-tools company.

API revenue is finally a business

The other shift Willison highlights is the API itself. For years, more API usage meant more losses — inference costs scaled faster than pricing, and every new enterprise customer made the unit economics worse.

That math has flipped. Token prices for equivalent capability have dropped roughly 10x in a year, while volume has exploded. Falling unit prices plus explosive growth in total usage is the same shape AWS rode in its early years: you’re not making money on any individual call, you’re making money because there are now an absurd number of calls.

The customer mix changed too. The paying base isn’t free-tier consumers — it’s companies wiring Claude and GPT into their actual workflows, and developers building products on top. That’s a far stickier and more defensible revenue base than ad-supported chat ever was.

So what is the consumer chatbot for?

If the money is in B2B and developer tools, the consumer apps start looking like something other than the main product. They look like marketing.

The free ChatGPT app and Claude.ai are how you build the brand, validate the model with millions of edge cases, and stay culturally legible. The revenue shows up downstream, in API contracts and enterprise seats. It’s the Google model — search is free, ads pay the bills — and the AWS model — invisible to consumers, enormous to the bottom line — converging into the same shape.

It also explains some recent behavior. The relentless push toward agents, IDE integrations, and “Claude for work” tiers makes more sense once you accept that the free chatbot is the funnel, not the destination.

What this means if you’re building

Two practical takeaways fall out of Willison’s diagnosis.

For anyone trying to compete with the frontier labs on general-purpose chat: don’t. That’s a marketing surface for companies with $10B+ in capital. The winnable PMF lives in specific workflows — legal review, code review, claims processing, anything where the customer can measure dollars saved per seat per month.

For individual developers: the $100–200 a month on Cursor or Claude Code looks expensive until you do the hourly math, at which point it’s one of the most lopsided trades in the industry. And every developer who pays it funds better coding models, which makes the tools sharper, which justifies the price. The flywheel is real.

If Willison is right, the “AI has no business model” critique is quietly running out of room. The more interesting question now isn’t whether AI pays — it’s who owns which vertical of PMF, and how deeply. In your own work, when was the moment AI stopped being a demo and started being something you’d actually pay for?