Heavy users of Claude Code - Anthropic's AI coding tool - regularly report burning through hundreds of millions of tokens per month. For context: one token is roughly three-quarters of a word, so 100 million tokens represents around 75 million words. Developers who plan architecture before prompting, keep sessions focused, and work deliberately across a few codebases typically land at 15-25 million tokens per month even with intensive daily use. The gap between that baseline and the power-user ceiling - sometimes north of 500 million tokens per month - comes down to a few specific habits.
The biggest driver is long context accumulation. Each message in a Claude Code session includes the full history of that conversation plus any files attached. A session that opens a 20,000-line codebase and runs for several hours can send millions of tokens with every single exchange - including short, one-line follow-ups. Developers who don't periodically prune attached context or start fresh sessions see costs compound fast. A second driver is pace: rapid prototyping with frequent restarts and loose prompts consumes far more tokens than deliberate, precise work, even if the actual output is similar in quality.
At current subscription tiers, developers who keep sessions lean stay comfortably within standard plan limits. The users hitting walls are almost always working with sprawling sessions attached to large codebases - not because they're producing more, but because every short message carries the full weight of a long conversation history.