Anthropic (economics)

Type: Company — inference-economics lens (unit economics only)

This is the economics-lensed stub for Anthropic in the inference-economics topic. The canonical Anthropic profile lives on the models rung at models/wiki/entities/anthropic.md; this page tracks only the (attributed, unverified) claims about Anthropic's inference unit economics, because the topic's central concept — the inference-margin question — turns on exactly these numbers.

Source culled — this claim is no longer evidence (flagged 2026-09-24). The podcast this claim comes from was removed from the KB in the 2026-08-06 cull, on the ruling that honest labelling of a conflicted source is not sufficient. It is not re-registered here, and re-registering it would reverse that ruling. Whether to strike the paragraph below or re-ground it on a source that still exists is an editorial call, open since 2026-08-20 — so the text is kept as the record of what was claimed, and nothing on this rung weighs it.

The claim, and its provenance caveat. On Podcast Alpha (2026-07-10), Dylan Patel of SemiAnalysis claimed Anthropic is FCF-positive, past $50B ARR, at 70%+ gross margin — well above Ben Evans' generic industry estimate of 40-50% inference gross margin. This is attributed / unverified, and carries a disclosed conflict of interest: SemiAnalysis is reportedly an Anthropic enterprise customer (~$7M/year on Claude Code), giving Patel motivated visibility but not independence. No primary Anthropic financial disclosure exists in this KB to check it against. Located web reporting from 2026-07-17 (unregistered; investor-deck lineage, unaudited) put Anthropic's present-period gross margin near 50%, with ~77% projected only for 2028 — see frontier.md.

Do not treat any of these figures as fact. They are logged here because the 30-point gap between "40-50% industry" and "70%+ Anthropic-specific" is the single number the topic most needs a primary source to resolve — not because the claim is established.

A second, structural reason to distrust the number (2026-07-22). Beyond the conflict of interest, the claim has no stated cost denominator — no utilization assumption, no cost unit, no as-of date. Patil's measurement (arXiv 2606.11690, 2026-06-10) shows effective cost on identical H100 hardware spanning $0.21-$15.25 per million output tokens with traffic shape alone. A gross-margin figure quoted without a utilization basis is not a precise claim that might be wrong; it is an imprecise claim that cannot be checked. The same objection applies with equal force to Evans' 40-50%. See the cost-measurement problem.

Key Contributions (to this topic)

• The margin claim states no cost denominator, so it is unfalsifiable as given — effective cost is a 36x-range quantity on fixed hardware. Evidence: moderate (measured ratios in Patil) (Patil)

• Cut the price of cached context, not the headline rate (2026-09-01). Claude Fable 5.1 kept $10/$50 per Mtok input/output and cut cache reads 75%, $1.00 → $0.25 — the deepest cache discount on Anthropic's list (1/40 of the input price; Opus 5 and Sonnet 5 sit at 1/10, Opus 5.5 at 1/20). Anthropic's own estimate: ~25% off typical workloads, up to ~45% off highly agentic ones. This is a pricing fact, not a margin fact: it says nothing about what a cached token costs Anthropic to serve. Evidence: strong for the prices, weak for the savings (vendor estimate) (Anthropic)

  Why it is on the rung: we rejected two pricing-policy items on 2026-09-13 as drift toward the retired Token Price Index. This one was admitted only because the cache-read line is itself the mechanism — the price of re-using stored model state — not a plain list-price move.

Mentioned In

• Token Pricing & the Inference-Margin Question — the Patel-vs-Evans margin tension; this stub holds the Anthropic-specific side of it.
• The Cost-Measurement Problem — why neither side of that tension is currently checkable.

Related Entities

• OpenAI (pricing) — the other frontier-lab economics stub in this topic (pricing behaviour, primary).
• vllm-cost-meter — the class of instrument that would be needed to check a claim like this one.

Changelog

• 2026-09-24 (source ingest) — Added the Fable 5.1 cache-read cut as the page's first primary Anthropic source, with the boundary reason it was admitted.
• 2026-09-24 — Flagged the culled source: the Patel podcast was removed from the KB on 2026-08-06 and this page still leaned on it. Moved it out of sources into culled_sources, kept the claim text as the record, and left the strike-or-re-ground decision open. Fixed the canonical-profile link (the old ai topic is now models).
• 2026-07-22 — Added the structural objection from Patil (2606.11690): the margin claim has no cost denominator and is unfalsifiable as stated — an objection that applies equally to Evans' 40-50%.
• 2026-07-16 — Stub created (economics lens). Holds the Patel-attributed Anthropic margin claim with the conflict disclosed. Canonical Anthropic profile stays at ai/wiki/entities/anthropic.md; this is a topic-lensed pointer, not a duplicate.