We analyzed 100 anonymized codebases from our beta cohort. We wanted to understand where token spend actually goes in real development workflows.
The results reshaped how we think about compression.
Where the tokens go
The biggest surprise was how much context gets repeated. Across the dataset, 41% of tokens sent to an AI API were exact or near-exact duplicates of content sent in a previous message within the same session.
The breakdown: - 22% repeated import blocks and type definitions - 11% commented-out code that developers had not yet deleted - 8% documentation strings on unchanged functions - 41% intentional context
That last number is the floor. Even with perfect compression, you are paying for that 41% of meaningful tokens. Everything else is recoverable.
Which languages waste the most
TypeScript and Python stood out as the most wasteful, for different reasons.
TypeScript projects tend to include large type definition files that get attached as context even when the relevant types are only a small fraction of the file.
Python projects suffer from verbose docstrings. Well-documented Python code is often 40-50% comments by token count.
Go and Rust projects were the leanest. Idiomatic code in these languages tends to be dense and explicit, with less ambient verbosity.
What this means for tooling
If you are paying attention to AI costs, the first thing to optimize is not your model choice. It is your context construction.
The model price differences between GPT-4o, Claude Sonnet, and Mistral Large are significant, but they are fixed per token. The number of tokens you send is entirely within your control.
Woozcode exists because this optimization is mechanical and tedious to do by hand. We have automated the parts that are safe to automate, and left the judgment calls to developers.