Willison argues that the AI labs have finally found PMF — not in chatbots, but in long-running coding agents that consume tokens at a scale that justifies the massive capex. He cites Anthropic's rumored first profitable quarter and anecdotes like a developer burning $2,180 of tokens against a $200 Claude Max subscription as evidence that the unit economics now work at industrial scale.
Does the arithmetic publicly: $5T-$10T of compute buildout across OpenAI, Anthropic, and hyperscalers needs to amortize over five years, implying $1T+/year in token spend. He argues an addressable market of ~200M knowledge workers and ~30M developers makes the numbers plausible if coding agents become the dominant workflow.
Argues that PMF for coding tools was actually reached sometime in 2024, well before this analysis, and that the more interesting and unsettled question is profitability. He says Willison's framing muddles two distinct claims, and the profitable-quarter rumor remains unverified.
Calls out the framing as 'AI psychosis' — the breathless tone of treating rumored profitability as a vindication of the entire capex thesis. Suggests the discourse is letting narrative outrun the actual financial evidence.
Argues tokens have no intrinsic cost or value — they're a unit defined by the labs themselves, so 'I used $2,180 of tokens' is circular reasoning. Compares it to a salesperson telling you you're getting a billion dollars of pots and pans for $19.99: the headline number only exists to flatter the purchase.
Raises the question every API customer should be asking: how do Anthropic and OpenAI retain users when something like GLM-5.1 ships open-weight at comparable quality for a fraction of the cost? Implies that current token economics depend on a moat that may not survive the next generation of open releases.
Simon Willison published an analysis on May 27 arguing that Anthropic and OpenAI have finally found product-market fit — not in chatbots, but in agentic coding tools that consume tokens at industrial scale. The post points to circulating rumors that Anthropic is on track for its first profitable quarter, and to anecdotes from companies stunned by their LLM bills. Willison cites one developer reporting $2,180.16 worth of tokens consumed against a $200 Claude Max subscription in a single month — a 10x usage-to-price ratio that Anthropic is willing to eat because the same workflow on metered API pricing would represent a paying power user.
The Hacker News thread (847 points at time of writing) is sharply divided. The core claim is not that LLMs are useful — that was settled in 2023 — but that a specific workflow (long-running coding agents) finally consumes enough tokens per knowledge worker to justify the capex. Commenter `trjordan` did the arithmetic publicly: between OpenAI, Anthropic, and the hyperscalers backing them, there is roughly $5T to $10T of compute buildout that needs to be amortized over the next five years, which implies more than $1T/year in token spend across an addressable market of ~200M knowledge workers and ~30M developers.
Not everyone is buying it. `aerhardt` argues the post conflates PMF (reached for coding sometime in 2024) with profitability (still unverified). `noddingham` calls out "AI psychosis" in the framing. `binary0010` raises the question everyone building on these APIs should be asking: how do the labs retain customers when GLM-5.1 ships open-weight at comparable quality for a fraction of the cost? And `prepend` lands the sharpest critique: "Tokens don't have an intrinsic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I'm getting a billion dollars worth of pots and pans for $19.99."
There are two coherent readings of this moment, and they lead to opposite conclusions about how to build.
The bull case: Coding agents are the first LLM workflow that scales token consumption linearly with developer productivity. A chat session uses maybe 10K tokens. A Claude Code session refactoring a service can burn 2-5M tokens in an afternoon. Multiply that by 30M developers and you get a market that justifies the buildout. Anthropic's reported margins on coding workloads are reportedly the inverse of their chat margins — chat is a loss leader, agents are the business. If true, this is the most important shift in the API economics since the GPT-3 launch: the unit of consumption is no longer the query, it's the agentic task.
The bear case: None of this is defensible. The Hacker News commenter who works in enterprise automation reports his largest clients are migrating to GPT-OSS 120B and similar open-weight models running on their own GPUs. The argument is straightforward: if your workload is high-volume agentic inference, the per-token economics collapse on dedicated hardware once you cross a usage threshold. Anthropic's $200/month subscription losing money on power users is not a feature — it's a tell that the pricing can't survive contact with workloads that actually use the product as intended. The $2,180-for-$200 anecdote is the same dynamic that killed unlimited mobile data plans.
The deeper issue is that "tokens consumed" is a vendor-defined metric. Prepend's pots-and-pans analogy is more cutting than it first appears. A 2M-token Claude session might do the work of a junior engineer for a day, or it might thrash through retries on a problem GPT-OSS would solve in 200K tokens. The industry has no agreed unit of useful work. Until benchmarks like SWE-bench Verified are tied to cost-per-resolved-issue rather than cost-per-token, we're flying blind on actual ROI.
Meanwhile, the open-weight curve is steep. GLM-5.1 (released Q1 2026) closed most of the gap on Claude Sonnet for coding tasks at roughly 1/8th the inference cost when self-hosted. Qwen3-Coder and DeepSeek-V4 are in the same zone. The pricing power Anthropic and OpenAI have today is real, but the half-life of that power is measured in quarters, not years.
If you're building on the Anthropic or OpenAI API, three things are worth doing this quarter.
First, instrument cost-per-outcome, not cost-per-token. Tag every agent run with the business task it completed (PR merged, ticket resolved, doc generated) and divide spend by outcomes. This is the only number that lets you compare Claude, GPT, and a self-hosted open model on equal footing. Without it, you're at the mercy of pricing changes that the labs will absolutely make once the subscription losses become a board-level conversation.
Second, build a swap layer for your inference provider now, while you still have the slack to do it. Most teams have already learned this lesson with cloud regions and databases. Yet production agent code is shot through with hard dependencies on Anthropic-specific tool-use formats, OpenAI-specific function calling, and prompt patterns that don't transfer. A thin abstraction over `messages.create` with a swappable backend costs a week and saves a quarter when pricing shifts.
Third, assume the $200 subscription tiers are subsidized and will be re-priced. If your business model depends on running Claude Code at 10x the subscription's intended usage, you have a clock running. The labs will either introduce hard rate limits, raise prices, or move power users to enterprise contracts. None of those outcomes are bad — they're just predictable, and worth modeling now.
Profitable quarters at Anthropic and OpenAI, if they materialize, are a milestone — not a moat. The same workloads that make these companies profitable also make their pricing maximally exposed to open-weight competition, because the customers spending the most are the ones with the strongest incentive to bring inference in-house. The next 18 months will reveal whether the closed labs can hold their lead through capability (better models, faster) or whether the agent stack commoditizes around open weights and the value migrates to tooling, evals, and integration. Bet your roadmap accordingly — and instrument the question so you can answer it with your own data, not theirs.
I feel like there's a bit of AI psychosis in this particular post.>"These are tools which burn vastly more tokens, but are also quickly becoming daily drivers for the work carried out by extremely well-compensated professionals.">"Somehow this fragment turned into headlines
I find this analysis confusing. PMF for coding was likely reached some time last year. Profitability, which is different, we don’t know. The article kind of confuses both without making a strong economic case or using numbers in a compelling way. I don’t understand what the Uber case has to do with
So how do openai and anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper?I don't see the business model working. My closest friend actually does automation software for large companies.He does not use Claude or openai at all. He primarily uses gpt 120
> $2,180.16 worth of tokens for $200“Tokens” don’t have an intrisic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I’m getting a billion dollars worth of pots and pans for $19.99.I think it’s funny how we are throwing critical thinkin
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a worl