A recent Forbes article claimed Anthropic's Claude Code Max plan costs users $5,000 per month in compute. This figure, however, stems from confusion between retail API prices and actual infrastructure costs. The $5,000 estimate reflects API-equivalent usage at Anthropic's $5/$25 per million token rates, but this is not what Anthropic spends to serve those tokens.

Looking at competitive open-weight models on OpenRouter reveals a stark contrast: Qwen 3.5 397B-A17B costs just $0.39/$2.34 per million tokens, roughly 10% of Anthropic's API rates. This pricing gap suggests Anthropic's actual compute costs for heavy users are around $500/month, not $4,800. While the $5,000 figure is accurate for Cursor's API bill, it highlights Cursor's higher costs for accessing Anthropic's models rather than Anthropic's own expenses.

The core issue is that developers prefer Anthropic's models despite the premium, not that inference is inherently unprofitable.