Developer questions massive token consumption claims
A Hacker News post asked why some developers burn through huge token quotas each month. The question drew four points and three comments. One commenter reported staying under one billion tokens monthly even during periods when AI handled most of the coding work. The gap between that experience and reports of full Claude or Codex 20x subscriptions suggests dramatically different workflows.
One coder stays under one billion tokens monthly through careful review
The commenter explained they cannot reach one billion tokens because they review everything the LLM writes. This manual gatekeeping caps consumption. They also built their own harness in C and a VSCode-like IDE in C++ using AI assistance. The C++ IDE felt slow which colored their experience. Even in months where the AI did most of the coding the total stayed below one billion tokens.
Custom tooling in C and C++ shapes token efficiency
Writing custom infrastructure changes how token budgets get spent. A C harness and a C++ IDE represent fixed investments that pay off over repeated use. The commenter noted the C++ IDE felt slow which may have discouraged iterative prompting. Slow feedback loops reduce the temptation to re-prompt. This contrasts with cloud-based assistants that respond instantly and encourage back-and-forth cycles.
Subagent avoidance keeps usage bounded
The commenter explicitly stated they never use subagents. Subagents multiply token consumption by delegating subtasks to fresh context windows. Each subagent invocation consumes its own input and output tokens. Avoiding this pattern eliminates a major source of runaway usage. Many agent frameworks enable subagents by default which can inflate monthly totals without the developer realizing.
What drives 20x subscription usage remains unclear
The commenter wondered what developers claiming full 20x Claude or Codex subscriptions are burning tokens on. Possible explanations include batch processing entire repositories generating tests documentation or refactoring pipelines. Some teams may run continuous integration loops that invoke models on every commit. Others might use recursive planning agents that iterate until tests pass. The Hacker News thread did not produce a definitive answer.
Token awareness separates sustainable from runaway workflows
Developers who track their token boundaries tend to build differently. The commenter who stays under one billion tokens does so through deliberate review and custom tooling. This awareness acts as a natural governor. As AI coding assistants become standard the gap between token-aware and token-blind workflows will widen. Organizations that measure and optimize token spend will gain cost advantages over those that treat quotas as unlimited.