Anthropic's recent blog post about AI coding capabilities masks a fundamental economic problem, according to a developer who spent four months experimenting with Claude Code. The author built a 40k-line application using Opus 4.6 but discovered that subscription pricing hides the true computational costs of serious LLM usage.

Agentic coding tasks that push models to their limits consume massive token volumes, making them far more expensive than advertised pricing suggests. At API rates, full utilization of a $100/month Claude Max plan would reportedly cost over $1000 monthly. Even single complex queries can burn through one million tokens, translating to roughly $75 per task.

These findings reveal a disconnect between consumer pricing and actual inference costs. While casual conversations remain cheap, intensive applications like code generation require exponential computational resources. The author warns that current subsidies make LLM coding appear viable, but this economic model cannot sustain itself indefinitely.

The experiment demonstrates both LLM coding's impressive capabilities and its prohibitive economics. Claude Code produced functional applications that would have been impossible for the author alone, yet the token consumption patterns suggest Anthropic is subsidizing user costs while developing newer models like Opus 4.7 to reduce expenses. This positions LLM coding as an economically unsustainable practice for most real-world applications.