Willison argues the rumored profitable quarter at Anthropic, combined with stories of companies shocked by their LLM bills and engineers happily burning thousands on Claude Code, is evidence that coding agents have found real PMF. The product has shifted from 'chatbot that helps you code' to autonomous worker billed against contractor budgets, which is why four-figure monthly bills per developer are tolerated.
The editorial endorses Willison's framing, emphasizing that coding agents crossed a threshold where engineering teams accept four-figure monthly bills because the productivity delta is legible. Revenue growth from ~$1B to $7B annualized at Anthropic in under 18 months is treated as the strongest signal that this is a real category, not a hype cycle.
Argues that product-market fit for coding agents was likely achieved last year and is not the interesting question anymore. Whether the underlying economics actually work — given inference costs, compute amortization, and the gap between revenue and capex — remains genuinely unknown.
Does the back-of-envelope math and concludes the industry needs roughly $5T–$10T returned over five years to justify current hardware investments. That implies $1T+ per year in token spending, a scale that current revenue trajectories — even at Anthropic and OpenAI's growth rates — fall far short of, suggesting eventual writedowns.
Points out that 'tokens' don't have intrinsic cost or value — quoting a $2,180.16 token bill is like trusting the vendor's own unit pricing to measure what was delivered. The framing inflates perceived productivity by letting the seller define the denominator.
Simon Willison's May 27 post landed a quiet bomb: Anthropic is rumored to be heading into its first profitable quarter, and the explanation isn't chatbots — it's coding agents burning tokens at rates that make legacy SaaS pricing look quaint. The post strings together a series of signals — companies expressing shock at their LLM bills, a widely-shared anecdote of a Claude Code user consuming $2,180.16 worth of tokens against a $200 subscription, and the broader observation that tools like Claude Code, Cursor, and Codex CLI have become daily drivers for well-paid engineers.
The numbers, where they exist, are stark. Anthropic's annualized revenue is reported to be in the $7B range, up from roughly $1B at the start of 2025. OpenAI is reported to be north of $13B annualized. The Hacker News commentariat is split on whether this is a milestone or a mirage. One commenter, [trjordan], does the back-of-envelope math: "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." Another, [aerhardt], pushes back: "PMF for coding was likely reached some time last year. Profitability, which is different, we don't know."
The most interesting community thread isn't the macro skepticism — it's the working assumption that coding agents have crossed a threshold where engineering teams will tolerate four-figure monthly bills per developer because the productivity delta is legible. The product is no longer "a chatbot that helps you code" — it's an autonomous worker that you bill against the same budget line as a contractor.
The Hacker News discussion makes the strongest case against Willison's framing, so let's steelman it before defending him. [prepend] notes: "'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." This is fair — list-price token math is marketing. But it misses the structural shift: the unit of work is no longer the API call, it's the agent session, and agent sessions consume tokens at orders of magnitude beyond what any chat-completion workload ever did. A single Claude Code task that touches twenty files, runs a test suite, and iterates on failures can easily consume hundreds of thousands of tokens. Multiply that by a team of twenty engineers running multiple sessions a day, and you arrive at the bill that's making CFOs sweat.
The open-source counter-argument is more serious. [binary0010] writes: "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." GLM-5.1, Qwen3-Coder, DeepSeek-V4 — the open-weight frontier has genuinely closed the gap on coding benchmarks. SWE-bench numbers from the Chinese labs are within a few points of Claude and GPT. If the model were the product, the moat would already be gone.
But the model isn't the product. The harness is. Claude Code's value isn't just Sonnet's reasoning — it's the file-editing semantics, the tool-use loop, the context-management heuristics, the way it knows when to read a file vs. grep, when to run tests vs. ask the user. Replicating that on top of an open-weight model is a serious engineering project. The labs that ship the harness have a six-to-twelve-month structural advantage even when the underlying weights commoditize. This is the same dynamic that kept AWS dominant after every primitive it offered was cloneable on bare metal.
The skeptical voice in the thread, [noddingham], calls the framing "AI psychosis" and questions whether the productivity claims actually translate to economic output. This is the honest open question. The revenue is real. The token consumption is real. What's not yet proven is that the output justifies the spend at the org level — whether a team paying $50K/year in Claude bills is shipping $50K more in value than they would have without it. Anthropic and OpenAI don't need to answer this question to be profitable. They just need enterprises to keep believing the answer is yes for another two years.
If you're shipping software with these tools, three things changed.
First, your AI line item is now a variable cost tied to agent autonomy, not a fixed seat cost. Budget accordingly. A team that adopts Claude Code aggressively can 5x its token spend in a quarter without onboarding a single new engineer. The old SaaS-budgeting heuristic of "$X per seat per month" no longer holds. Treat it like compute: monitor usage, set per-developer ceilings, and build cost-awareness into the development workflow itself.
Second, the open-source escape hatch is real but it isn't free. Yes, GLM-5.1 or Qwen3-Coder on a local H100 or a cheap inference provider will cut your per-token cost by 10-50x. But you'll spend the savings — and then some — building the harness, the eval loop, the security review, and the integration plumbing that Claude Code and Cursor give you out of the box. The right time to make that switch is when your token bill exceeds the loaded cost of an engineer dedicated to maintaining the alternative. That's a higher bar than it sounds.
Third, the architectural assumption to revisit is whether your codebase is *agent-ready*. Agents perform dramatically better on codebases with clear module boundaries, strong tests, good naming, and documented entry points. The teams getting the most leverage out of these tools aren't the ones with the best prompts — they're the ones whose codebases are legible to a stochastic reader. If your repo is a tangle of implicit conventions and tribal knowledge, the agent will burn tokens flailing, and your bill will be high without the corresponding output.
Willison's piece will be wrong if the open-source frontier closes faster than the harness lead, or if the productivity claims fail to convert to retained enterprise revenue when the next budget cycle hits. Both are plausible. But the immediate read is that the two labs have stumbled into a pricing model — pay-per-autonomy — that scales with how aggressively customers adopt agents, and customers are adopting aggressively. The next eighteen months will tell us whether this is the Uber-of-tokens moment (real category, eventually profitable, painful path) or the WeWork-of-tokens moment (real demand, broken unit economics). Bet your stack accordingly.
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