Look at this. The 151,000 token usage is when I have Agent mode on with 5 files in context.
Turn off agent mode for Manual mode and it’s suddenly 30,000 tokens.
All the costs are agent mode - and their edit_tool calls fails for me 50% of the time, or the agent randomly uses grep on the codebase instead of reading what’s inside the chat context.
All these tool calls, and their agent is infantile. It’s definitely a regression. Claude 3.7 used to do like 50 tool calls where gemini 2.5 pro would immitate the same code change in a single tool call.
The problem with the crazy token usage is the tool calls and the agent mode. Just turn it off, and manually apply code changes and the surge pricing drops. Agentic coding today just ■■■■■.
