Anthropic’s token prices are falling faster than its compute bill is rising 🪙

Anthropic said revenue growth from increased usage was "partially offset by a decrease of approximately $600 million from price reductions introduced as part of new model versions and contractual discounts," according to its amended SEC filing. The developer of the Claude family of AI models has filed for an initial public offering, announcing on Monday that it plans to list its Class A common stock on the Nasdaq Global Select Mark.

Compute and infrastructure expenses surged from $5.56 billion to $9.86 billion, an increase of roughly 77%. Anthropic spent roughly 58 cents on compute and infrastructure for every dollar of revenue, versus 48 cents a quarter earlier. The measure includes computing used for research, training, and internal operations—not just serving paying Claude customers—so it is not a per-prompt profit margin.

The company’s price reductions reflect both lower prices and efficiency gains rather than straight discounts across every task. Claude Opus 5.5, which lowered token prices compared with Opus 5, was recently released and is more capable than any previous model, sometimes reaching parity with Fable. Anthropic also introduced Haiku 5.5 in October as a lower-cost model for high-volume work. The company says Fable 5.1 reduced the cost of typical workloads by around 25%, while Opus 5.5 brought customers’ cost per task down roughly 40% compared with Opus 5.

Anthropic typically reserves compute capacity 12 to 18 months before it becomes available, leaving much of that spending relatively fixed in the near term. The quarter included $5 billion of compute under short-term commitments that Anthropic says cost significantly more than comparable capacity secured through multiyear contracts. The expense helped push the company from a $974 million operating profit in the second quarter to a $567 million operating loss in the third.

Anthropic’s updated disclosures put aggregate lease and purchasing commitments at roughly $895.4 billion, substantially all related to compute and infrastructure. The company also said users have historically expanded consumption after releases of cheaper, more capable AI, while its filings acknowledge the difficulty of matching demand with available supply.