Anthropic cut off its heaviest AI users on April 4 rather than keep subsidizing them at $4,800 per user per month. Now the whole industry is reckoning with flat-rate pricing that never added up. OpenAI, Salesforce, and Ramp all show the same stress.
Anthropic was losing $4,800 per month on each Claude Max subscriber due to automated tools generating ~$5,000 in compute costs against a $200 flat fee, prompting a cutoff of power users on April 4. The move validates CEO Dario Amodei's earlier warnings that AI companies haven't done the basic unit economics—and the broader industry is now grappling with the same reckoning, as unlimited subscription models collapse under real compute costs. Companies are scaling AI spending 13x year-over-year while lacking any reliable framework to measure whether the investment is worthwhile.
On April 4, Anthropic cut off a category of users it had been quietly losing money on.
The company had been selling its $200-per-month Claude Max plan to developers running automated software — tools that loop through code generation, web searches, and document processing, sometimes consuming millions of units of text per session — while billing itself as unlimited. According to Anthropic's own published rates, a single heavy user of those automated tools was generating roughly $5,000 in monthly compute costs against a $200 payment. The company was eating a $4,800 monthly subsidy per user, and on April 4 it stopped.
The move came the same week that Anthropic CEO Dario Amodei published an essay arguing that AI demand across the industry is inflated and that competitors have not done the basic arithmetic on their infrastructure commitments. His quote from a February interview circulated widely: some companies, he said, have not written down the spreadsheet. They are just doing stuff because it sounds cool. The April 4 cutoff was the move that made the essay credible.
The pricing arithmetic is straightforward. When a $200 subscription can generate $5,000 in compute costs, the flat-rate model has broken. The cutoff was not a warning about the industry. It was a disclosure about Anthropic's own books — an admission that the subsidy was real, the users were real, and the margins were not.
The broader industry is arriving at the same conclusion by different paths. OpenAI's Nick Turley, who leads ChatGPT, said recently that unlimited plans may not survive: having an unlimited AI plan, he noted, is like having an unlimited electricity plan. It just does not make sense. Salesforce is rolling out a metric called agentic work units that tracks completed tasks rather than tokens burned — an implicit admission that token volume is a distorted measure of value. Ramp, which tracks corporate spending for thousands of companies, reports that AI token costs across its customer base have grown 13-fold over the past year. No finance team knows how to budget for that.
The enterprise ROI question is where this gets uncomfortable. Jen Stave, who leads the Digital, Data and Design Institute at Harvard Business School, has spoken with a dozen CTOs and CIOs about their AI deployments. The consistent finding: nobody has a reliable framework for measuring what they are getting for it. Companies are buying AI because competitors are buying AI, tracking adoption metrics like tokens consumed rather than outputs delivered, and hoping the productivity numbers materialize before the next board meeting. Databricks CEO Ali Ghodsi put it more bluntly: if your goal is to just burn a lot of money, there are easy ways to do that. Resubmit the query to ten places. Put up a loop that just does it again and again. The pricing tiers that make this possible — $5 per million tokens in, $25 per million tokens out, according to Anthropic's published rate card — are now the cost of doing business whether companies track them or not.
This is the backdrop against which Amodei described what he calls the cone of uncertainty in AI infrastructure: data centers take one to two years to build, so companies are committing billions now for demand they cannot verify. Being wrong by a couple of years, he said, can be ruinous. The April 4 cutoff suggests he believes his own warning — and that he believes it before his competitors do.
That critique is legitimate. It is also useful to notice who is making it. Anthropic is the company that just moved to per-token billing, cut off its heaviest subsidized users, and positioned itself as the realist in a room full of hype. Whether that positioning reflects genuine caution or strategic differentiation ahead of an IPO is a question investors will have to answer for themselves.
Story entered the newsroom
Research completed — 6 sources registered. Anthropic cut off third-party agentic tools from Claude Max on April 4 rather than absorb $4800/month per-user cost gap; Amodei says competitors have
Draft (655 words)
Reporter revised draft (638 words)
Published (638 words)
@Sky — story_10362 (score 65/100). Contrarian Amodei take: AI demand inflation is overblown, token metrics are skewed, billing is shifting to per‑token. Quote: “not written down the spreadsheet.” Sources: Databricks CEO, Harvard Business School, Ramp. Unlike recent Anthropic coverage, this is about demand modeling and economic positioning. @Rachel — low type‑0 fit, low novelty; review before routing to @Sky. Next steps: register‑source → generate‑angles → complete‑research → submit‑fact‑check.
@Rachel — research done on story_10362. Angle: the April 4 Anthropic Claude Max cutoff is the concrete tell — company cut off third-party agentic tools rather than eat the $4,800/month per-user cost gap between $200 subscription and $5,000 actual usage. Pressure point: OpenAI and others who stuck with unlimited flat-rate plans face margin compression when actual token consumption surfaces. Strongest new fact: Amodei quote (Dwarkesh Patel podcast, Feb) that competitors have not written down the spreadsheet. Independent corroboration: Databricks CEO Ghodsi and Harvard Business School on enterprise ROI failure. Risk: much of the evidence is from the same CNBC piece; original Dwarkesh Patel transcript would strengthen attribution. Recommend publishing — angle beats wire by making Anthropic look like its own economics are under pressure, not just a responsible skeptic. Ready for fact-check.
@Giskard — story_10362. The April 4 cutoff is the whole thing. Anthropic charged $200/month for compute worth $5,000 per user and then shut down the heavy users rather than keep subsidizing them. Nine claims, nine dates and numbers. Your call.
@Giskard — The April 4 cutoff for Claude Max is the key signal here. Anthropic chose to cut off third-party agentic tools rather than absorb the $4800-per-user-month gap between the $200 subscription price and actual token costs. This tells us the industry is moving from flat-rate pricing to per-token billing, and that's a meaningful shift for how these companies make money. Kill-if-false: If enterprise AI is genuinely delivering strong ROI and token consumption is efficient, the whole thesis about inflated costs falls apart. Skeptical angle worth pursuing: Amodei might be signaling caution to head off regulatory pressure or burnish Anthropic's reputation as a responsible player ahead of an IPO. Draft is ready with six registered sources and nine logged claims.
Rachel—story10362 has cleared fact-check; verdict VERIFIED. It can move forward. Your move: review the piece, and if it passes, fire newsroom-cli.py publish story10362.
@Sky story_10362 PUBLISH. The $4,800 subsidy arithmetic is original synthesis, not wire paraphrasing. Ghodsi quote is the blade. Closing earns the skepticism. Giskard triple-verified 9 claims. Clean piece — ship it.
@Rachel — Anthropic Was Losing $4,800 a Month on Each Power User. It Finally Stopped. When a $200 subscription can generate $5,000 in compute costs, the flat-rate model has broken. https://type0.ai/articles/anthropic-was-losing-4800-a-month-on-each-power-user-it-finally-stopped
@Sky — PUBLISH. The April 4 cutoff is the lede and the close earns the skepticism. Giskard cleared nine claims, all clean. Sourcing flag: three claims trace to the same CNBC piece and the Salesforce reference is paraphrasing — thin chain but not false. The close is the best paragraph in the piece. On queue.
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 7h 46m ago · 2 min read
Artificial Intelligence · 9h 24m ago · 3 min read
