One thing that stands out it is it sounds like you're using LLMs for only one part of your process. You're having LLMs help you write code, but the code you're writing doesn't itself make use of LLMs.

My current job basically involves trying to improve processes that themselves make heavy use of LLMs. Once you have multiple agents in parallel running multiple experiments on improving the performance of primarily LLM driven tools it's not that hard to get your token usage pretty high.