Anthropic and OpenAI Have Found Product-Market Fit, Says Simon Willison

Independent AI researcher and longtime open-source maintainer Simon Willison published a short, widely circulated post this morning arguing that Anthropic and OpenAI have finally hit real product-market fit — and that the product isn't ChatGPT or Claude.app. It's the coding agents: Claude Code (and its consumer-facing sibling Cowork) and OpenAI's Codex.

The post hit the front page of Hacker News within hours and is being passed around by AI investors, founders, and infrastructure teams as a clean summary of why the two leading labs suddenly look like profitable software businesses.

Willison's argument in one paragraph

The thesis: for years, frontier labs were giving away their best models at heavily discounted enterprise prices, hoping to lock in workflows. That changed in April. Both companies raised prices on their newest frontier models, both moved their largest enterprise customers onto list API rates instead of discount tiers, and both are now seeing real demand at those prices — particularly for coding agents.

GPT-5.5, released April 23, is about 2x the API price of GPT-5.4. Opus 4.7, released April 16, runs roughly 1.4x the price of Opus 4.6 after accounting for the new tokenizer. Pricing power, in other words, is now real.

Why coding agents specifically

Willison's case is that general-purpose agents — agents that can plan, write code, edit files, run terminals, and use tools — are the first AI product where the value to the customer clearly exceeds the cost of the tokens. A senior engineer's time is expensive. Spending tens of dollars in inference to save them an afternoon is a trivial trade.

Both companies have built around this. Claude Code ships as a CLI and IDE integration that runs autonomously in the developer's terminal. Codex is OpenAI's equivalent — an agent that takes tasks and works them through to completion. Both companies have layered enterprise controls, audit logs, and security policies on top.

The result is that the same companies that pushed back on $20-per-seat ChatGPT licenses are now signing seven- and eight-figure deals for coding agent capacity.

The supporting evidence

Willison points to several pieces of data outside his own analysis:

Anthropic disclosed that it's now spending roughly $1.25 billion per month with a single compute vendor for extra inference capacity. That's a number you only sign up for if your inference is generating revenue at scale.

Microsoft reportedly canceled some Claude Code licenses for financial reasons tied to its fiscal year end — a sign that even Microsoft, OpenAI's largest patron, was paying real money for Anthropic's coding product.

Uber publicly reported budget overruns on its AI coding rollout. Companies don't overspend on tools nobody uses.

And both companies are now planning IPOs. Anthropic just closed a $30 billion-plus funding round at a $900 billion-plus valuation, and OpenAI is reportedly targeting an IPO later this year at $25 billion ARR.

Why this matters

The post matters because it's the cleanest argument yet that the AI industry has crossed from "expensive science project" into "profitable category." For the last two years, skeptics have pointed to OpenAI's losses, Anthropic's compute bills, and the lack of clear daily-use cases as evidence that frontier labs were burning cash on a bet that hadn't paid off.

The coding-agent thesis cuts through that. If two companies can charge enterprise prices for their best models and customers will keep buying because the product is paying for itself in saved engineering hours, then the business model works. The compute bills become a cost of goods sold, not a runway-burning subsidy.

The counterargument

The Hacker News discussion under Willison's post surfaces the obvious risks. Open-weights models from DeepSeek, Qwen, GLM, and others are closing the gap on coding benchmarks, and several enterprises are starting to deploy them inside their own infrastructure. Cheap, good-enough open models could put a ceiling on what Anthropic and OpenAI can charge.

There's also a CNBC piece from last week arguing that "cheap AI could derail OpenAI and Anthropic's IPOs" — making essentially the inverse case. If the value of the agent product is in the orchestration and not the underlying model, the moat may be shallower than the price hikes suggest.

What's next

Willison's analysis lands in the middle of a bigger debate about whether AI is "the bubble" or "the foundation." His take won't settle the question, but it does give shape to one piece of it: enterprises are actually paying enterprise prices for coding agents, and that's something they weren't doing 12 months ago.

For founders building on top of frontier APIs, the practical read is that pricing is unlikely to fall as fast as it has in past cycles. Plan capacity and unit economics around list prices, not discounted rates.

Bottom line

Product-market fit in AI has been declared so many times it's become a punchline. Willison's piece is notable because it points to a specific product (coding agents), a specific signal (price increases that stick), and a specific buyer behavior (enterprises paying list). If he's right, the next 12 months in AI look less like a search for the killer app and more like a build-out of the one that just arrived.