A viral Forbes claim that Anthropic loses $5,000 per Claude Code Max subscriber is debunked by distinguishing retail API prices from actual compute costs. Using OpenRouter pricing for comparable open-weight models like Qwen 3.5 397B and Kimi K2.5 as a benchmark, actual inference costs are roughly 10x lower than API list prices. This means the heaviest Claude Code users cost Anthropic around $500/month in real compute, not $5,000 — a $300 loss at most, not $4,800. The $5,000 figure likely originates from Cursor's internal analysis, since Cursor pays near-retail API prices to Anthropic. For the average user spending ~$6/day in API-equivalent tokens, Anthropic's actual cost is around $18/month against a $20–$200 subscription, making per-user inference likely profitable. The broader 'AI inference is a money pit' narrative is characterized as misinformation that benefits frontier labs by obscuring their large markups.
Table of contents
What the $5,000 figure actually isThe OpenRouter reality checkSo who is actually losing $5,000?Anthropic is not a profitable company. But inference isn't why.Sort: