Anthropic Cut Its Flagship 20% the Day Before the Roadshow — Its $895 Billion Compute Bill Shows Who Actually Wins

Anthropic cut flagship input and output prices 20% below GPT-6 Astra at $8 and $40 per million tokens against $10 and $50, while the 80% gap applies only to cached reads, where Anthropic charges $0.20 per million against OpenAI's $1.00. For workloads that re-send a long prefix thousands of times, that cached rate is the number that actually drives a bill, and it is where Anthropic's price advantage is decisive rather than cosmetic.

Anthropic is not uniformly cheap. It charges $16 per million for a one-hour cache write against OpenAI's $12.50, pricing the durable write at twice its own input rate, while undercutting the five-minute write at $10. The pricing tells you exactly where the company wants volume: repeatable, long-context workloads that read a cached prefix many times, which is the shape of coding agents and not the shape of a single chat query.

Anthropic's first-half net loss of $80.7 billion reads like a cash crisis and is mostly an accounting consequence of the company's own success. $72.4 billion came from revaluing convertible notes, derivative liabilities and SAFEs higher as Anthropic's enterprise value rose, leaving an operating loss of just $693 million on $16,286 million of revenue. Any investor anchoring on the headline loss is anchoring on a mark, not a burn.

That structure makes the IPO a balance-sheet event as much as an earnings event. The notes convert into Anthropic equity at listing — Amazon's convert into a stake capped at 12% and could create as many as 1.84 billion shares — while Google's $20 billion of notes plus an option for up to $10 billion of Class N stock at the IPO price put a second bidder on the register. An IPO converts roughly $68 billion of supplier-financed compute obligations into equity, which is why the reported loss looks catastrophic while the funding position looks strong.

A 20% price cut requires 25% more tokens to hold revenue flat, and Anthropic is growing far faster than that. Third-quarter revenue rose 1,127% year over year and nine-month revenue reached $33.2 billion, so Anthropic can absorb a 20% discount and still print higher revenue. That is why this reads as a volume-for-margin trade rather than a defensive cut.

The derived numbers sharpen the point. Nine-month revenue of $33.2 billion less first-half revenue of $16,286 million implies third-quarter revenue of roughly $16.9 billion, a $67.7 billion annualized run-rate that is roughly level with the $70 billion OpenAI expects to reach by the end of 2026. Two labs, two private marks, and almost the same current revenue — which is why a price cut from one of them is a direct mark-down on the other's implied per-token value.

The bill for the price war is not paid by the labs. Anthropic has $895.4 billion of contracted compute and infrastructure commitments, with $447.5 billion due before Dec. 31, 2030 and $69.8 billion falling due within a year of Sept. 30, 2026. Contracted capacity does not shrink when the price per token falls, so deflation converts straight into orders for accelerators, foundry wafers and high-bandwidth memory.

The margins confirm where value is actually being captured. In the same quarter Anthropic cut prices, SK Hynix reported KRW 104.6 trillion of revenue and KRW 83.3 trillion of operating profit, a 79.7% operating margin, with revenue up 328% year over year. NVIDIA posted $96.2 billion of revenue in its fiscal second quarter ended July 26, 2026 at a 75.0% gross margin, and Broadcom turned $29.6 billion of revenue into $20.5 billion of gross profit in its fiscal third quarter ended Aug. 2, 2026.

On the other side of the same price sheet, buyers of tokens get cheaper. Nearly half of Anthropic's revenue — 48% in the first half of 2026, up from 11% in 2023 — reaches customers through the cloud marketplaces of Amazon, Google and Microsoft, and the top five customers now account for about 25% of revenue with none above 10%. The nine-month figure slipped to 44%, implying roughly 40% of third-quarter revenue came through cloud marketplaces, so Anthropic is reducing its dependence on the cloud channel just as it prices against the rival.

Marketing could begin as early as Nov. 10, fifteen days after the prospectus went public on Oct. 26, and as of Nov. 10 the S-1/A still shows a blank price range and share count. The Wall Street Journal reported on Nov. 2 that investors are pressing for a $1.5 trillion to $1.6 trillion valuation, while reporting on Nov. 8 suggested a postponement remains possible.

In days to quarters, three things move first: any matching OpenAI price response, the Anthropic price range, and the fourth-quarter report that will show whether net revenue retention holds near 900%. A rival cut inside the roadshow would reset the price grid within weeks and put the $190-200 billion 2028 projection under immediate review.

Over one to three years, the commitment book is what matters. About $447.9 billion of the $895.4 billion falls due between 2031 and 2045, roughly $30 billion a year against projected2028 revenue of $190-200 billion, a structural cost floor of about 15% of revenue that token price deflation does not touch. Anthropic wins if falling prices pull enough volume to fill that capacity; it loses if it has to keep discounting to fill it.

The honest read is that Anthropic is not cutting prices because it is losing; it is cutting because it can afford to, and because a cheaper flagship is the most persuasive page in a prospectus that must support a $1.5 trillion valuation on revenue that does not exist yet. Falling token prices are good for the buyers and good for the chipmakers; they are only good for the labs if volume fills capacity already signed for. The roadshow will tell you which side of that trade the market believes.