Anthropic cuts Claude prices as its compute bill soars: the cost of leading becomes visible in the S-1

By Yahoo Finance · November 3, 2026.

Anthropic is cutting Claude's prices while its compute bill grows much faster than its revenue. According to its amended filing with the SEC (the US Securities and Exchange Commission), revenue growth from higher usage was “partially offset” by a reduction of about $600 million, attributed to price cuts on new model versions and to contractual discounts.

Spending on compute and infrastructure rose from $5,560 million to $9,860 million, an increase of around 77%. Put proportionally, Anthropic spent about 58 cents on compute for every dollar of revenue, compared with 48 cents a quarter earlier. A caveat on how to read this: the figure includes research, training and internal operations, not just serving paying customers, so it is not a per-query margin.

Nor is it $600 million “lost.” It is the net effect of lower prices on growth that, by volume, remains strong.

Operating results swung sharply: from a $974 million profit in the second quarter to a $567 million loss in the third. According to the article, a factor was $5,000 million in short-term contracted compute, which, according to the company, costs it considerably more than capacity secured through multi-year contracts. Anthropic typically reserves capacity 12 to 18 months in advance, so much of its spending is rigid in the short term.

The cuts are real. Claude Opus 5.5 lowered per-token prices (the unit in which usage is billed) compared with Opus 5, and Haiku 5.5, launched in October, is positioned as a much cheaper model for high volume. According to Anthropic, Fable 5.1 reduced the cost of typical workloads by around 25%, and Opus 5.5 cut the cost per task by nearly 40% compared with Opus 5; both figures combine lower prices and efficiency gains.

Aggregate lease and purchase commitments, almost all for compute, stand at about $895,400 million. Reuters has also reported doubts about whether infrastructure partners will be able to finance all that capacity.

**Our take.** This is not a crisis; it is the arithmetic of a strategy: make it cheaper so customers consume more. Anthropic maintains that, historically, usage expands after each price cut. It works if demand grows faster than prices fall; the problem is that compute supply is bought 12 to 18 months in advance, and demand gives no warning. One quarter of operating losses does not invalidate the thesis, but it shows how narrow the room for maneuver is.

In general, those selling the “shovels” (chips, cloud, energy) get paid sooner and with less risk than the labs. The article itself points to Google and Nvidia as major beneficiaries of Anthropic's growth. And competition limits the easy way out: according to the source, OpenAI, which is also preparing to go public, offers a competitive model with more generous usage quotas, which curbs any price increase.

For readers, the lesson is twofold. In the short term, cheap AI is good news for businesses and developers, but it is worth keeping an eye on the sustainability of whoever provides it, especially if quotas run out quickly. In the long term, the underlying trend is what interests us: cost per task is falling, and that deflation is the path toward an abundance of cognitive capacity accessible to everyone. The question the numbers leave open is who finances the transition, not whether the destination is the right one.