AI coding costs rise due to inefficient token consumption

Research and industry guidance highlight a growing challenge in AI-assisted development: the unexpected rise in costs due to inefficient token consumption. While generative AI and agents improve coding speed, companies are facing budget overruns caused by high token usage.

A study by the Fiddler research team compared Anthropic’s Claude Haiku and Claude Sonnet models to analyze the relationship between token unit prices and actual repair costs. Despite Haiku having a lower unit price, the study found it could be significantly more expensive to use for specific tasks. In experiments involving Python bug fixes, the cost per successful task with Haiku was 8.4 to 14.1 times higher than with Sonnet.

The discrepancy was driven by Haiku generating excessively long responses. The model often predicted and wrote out numerous subsequent commands and hypothetical execution results instead of waiting for actual feedback. This behavior increased both output costs and subsequent input costs, as the entire conversation history—including the redundant text—had to be re-processed in each turn.

To address these issues, industry tools including Claude Code, VS Code, GitHub Copilot, and Cursor have shared best practices to prevent “token waste.” Effective strategies focus on reducing unnecessary output and managing session history to ensure that high-quality code generation remains economically viable for large-scale development.