Five thousand dollars per user, per month. That's what a Forbes article claimed Anthropic spends on each Claude Code Max subscriber. The number went viral. It also appears to be wildly wrong.
The $5,000 figure comes from multiplying heavy usage patterns against Anthropic's retail API prices - $5 per million input tokens and $25 per million output tokens for Opus 4.6. At those rates, a power user burning through 150-200 million tokens per day with a 95% cache hit rate could indeed rack up $5,000 in API-equivalent charges against a $200/month subscription.
But retail API prices are not the same as compute costs. Not even close.
The 10x Gap Between Price and Cost
Competing models of similar capability tell a different story. On OpenRouter, Qwen 3.5 397B charges $0.39 per million input tokens. Kimi K2.5 runs at $0.45 per million input. These providers are profitable at those prices. That puts them at roughly one-tenth of Anthropic's retail rates.
If Anthropic's actual inference costs follow a similar ratio, the math changes dramatically. That scary $5,000 power user actually costs Anthropic around $500 in compute. The real loss on an extreme subscriber drops from $4,800 to roughly $300 per month.
And extreme users are rare. According to the analysis, 90% of Claude Code subscribers consume under $12 per day in API-equivalent spend. The average user sits around $6 per day, or about $180 per month at retail rates. At actual compute costs, that translates to roughly $18 per month - well under the $200 subscription price.
Who Actually Pays $5,000
The $5,000 figure does describe something real - it's approximately what a company like Cursor would pay Anthropic at retail API rates for a heavy user. It represents the revenue Anthropic forgoes by offering flat-rate subscriptions instead of per-token billing. That's a pricing strategy decision, not a cost crisis.
Fewer than 5% of Max subscribers hit the weekly rate caps, which suggests the vast majority of users aren't anywhere near the theoretical maximum.
The Real Money Pit
Inference costs, even at scale, aren't what keeps Anthropic burning cash. Training runs, research staff, and GPU procurement dwarf the per-user serving costs. The $5,000 claim makes for a dramatic headline, but it confuses the price Anthropic charges others with the price Anthropic pays itself. On the average Max subscriber, Anthropic is likely turning a profit.