Anthropic's IPO rests on the very price it just cut

A token of Claude's work carried two different numbers five weeks apart. In mid-September, Anthropic collected about $1.70 per token on average; by late October it was $0.83 — the same work, priced at a bit under half ^[1]. Part of the fall was customers steering their jobs toward cheaper models ^[2]; the rest was the price list itself, and Anthropic is the one that marked it down. On September 22 the company cut its API prices 20% — the rate businesses pay to call the model — and its cache-read prices 60%, cache-read being the discounted rate for reusing data the model has already seen, and framed the move as handing its own efficiency gains to customers ^[3]. OpenAI matched within ninety minutes.

passing these efficiency savings on to our customers in the form of price cuts and rate limit increases. — Anthropic

That price list is now the unit of account in the document Anthropic filed to go public — the S-1, dated October 29. Its centerpiece exhibit reports that the retail value of the company's internal Claude usage, priced at what it charges outside customers, passed its monthly cash payroll in April. It is not spending more cash on AI than on salaries; the claim is that Claude was doing work worth more than the salaries cost, by April, at list price ^[4]. The metric stands on the very prices the war was cutting. Every time Anthropic lowers its list to pass efficiency gains along, the same internal usage is worth less on paper, so the headline number moves down against the company's own wins. The retail tag also prices a pattern its customers don't much buy: the filing touts that Claude writes more than 90% of Anthropic's new code ^[4], while the company's own research on its Australian users finds they lean toward back-office work over coding ^[5]. The warrant for pricing a token against a wage is the founder's own. Two weeks before the filing, Dario Amodei put the thesis in one line.

AI isn’t a substitute for specific human jobs but rather a general labor substitute for humans. — Dario Amodei

If AI is a general substitute for labor, a company whose model does most of its own work is the leading edge of the post-labor economy — and the value of that work, marked at list price, is the number to show. The sector is deep in prove-it-pays season: Santander booked €194 million of "business value" from AI in nine months ^[6]. The sharpest outside attack on Anthropic's metric landed the day before the filing. BCA Research named a hidden cost — the "AI validation tax," the hours employees spend checking AI-generated work — that sits on top of the token bill; and it warned separately that rising token counts may reflect a reasoning model's extra rounds of computation rather than more checked output. Priced at retail, the exhibit measures consumption, not verified work ^[7]. While the filing was pitching that number up, the rest of the week was pricing the other side of the same variable. None of these instruments is priced in the exhibit's unit; each is a bet on what a token of AI work is worth, and each was taken short. Jump Trading opened an over-the-counter desk — trades struck directly between two parties, off the exchanges — selling derivatives that let chip owners and data centers insure against the rental rates on their compute falling ^[8].

You don’t need to be bearish to want some insurance. — Jump Trading

The cost of insuring the hyperscalers themselves — against their stock or their debt falling — is at its highest since 2021 ^[8]. OpenAI asked for $30 billion, cut the ask to $20 billion, and collected $11 billion ^[9]. Accenture capped what its staff can spend on Claude at $29 a head per month and pulled access from the heaviest users ^[10]. The buyers were in the same trade, on the other side. Exchange-traded funds have been stacking pre-IPO Anthropic shares through special purpose vehicles — funds set up solely to hold the shares — with one putting roughly 9% of its own portfolio into them and another about 2%, deliberately copying the SpaceX playbook ^[11]. That precedent is fresh: SpaceX debuted at $150, ran to $225, fell under $100, and has recovered to $144 — below its own debut within weeks ^[12]. The one honest limit on the short side is BlackRock, which reported the week before the filing that AI compute now pays back its infrastructure in under three years, down from four ^[13]. Then there is the last item in the document, from the people who wrote it. Anthropic's seven cofounders have pledged 80% of their personal stock to a fund meant to cushion what the filing calls the "economic and societal fallout" of artificial intelligence. The mechanism adds 2% of the company's equity to that fund for every 50% the share price rises above the IPO price ^[14]. The same share price that will test the internal-usage exhibit is the number that scales the payout into the fallout fund. One cover, one number, two positions on it.