No, it doesn't cost Anthropic $5k per Claude Code user
Viral $5k loss claim confuses retail API prices with actual 10x cheaper compute costs.
A viral narrative claiming Anthropic hemorrhages $5,000 per user on its Claude Code Max subscription has been debunked. The figure originates from a Forbes article citing Cursor's internal analysis, but it confuses Anthropic's retail API pricing with its actual compute costs. At Anthropic's listed rates of $5 per million input tokens and $25 per million output tokens for Opus 4.6, a power user could indeed rack up a $5,000 API bill. However, this is not what it costs Anthropic to serve those tokens.
Comparing to competitive pricing on platforms like OpenRouter reveals the real cost structure. Models comparable to Opus 4.6, such as Qwen 3.5 397B or Kimi K2.5, are served for about $0.39-$0.45 per million input tokens—roughly 10x cheaper than Anthropic's retail API prices. These providers cover compute, GPU costs, and turn a profit, indicating the actual inference cost for a heavy Claude Code user is closer to $500. This results in a $300 monthly loss on the most extreme users, not $4,800, and Anthropic states fewer than 5% of subscribers hit such caps.
The real financial pressure is on third-party platforms like Cursor, which must pay near-retail API prices to integrate Claude, making their cost for a power user approximately $5,000. For Anthropic, the massive expenses lie in training frontier models and researcher salaries, not inference. The 'AI inference is a money pit' narrative benefits frontier labs by justifying high API markups, while the actual per-token serving cost for an average subscriber is likely profitable.
- The $5,000 figure uses Anthropic's retail API prices ($5/$25 per MTok), not the ~10x cheaper actual compute cost seen on OpenRouter.
- Real inference cost for a heavy user is ~$500, leading to a $300 loss, not the viral $4,800 loss claim.
- Third-party tools like Cursor face the $5k cost, but for Anthropic, inference is not the primary drain; model training and R&D are.
Why It Matters
Clarifies the real economics of AI services and challenges narratives that justify high API pricing markups for end-users.