I recently used the Agent mode (with fable 5) to fix a rule bug while contributing to a opensource project. The resulting fix was literally just 3 lines of code. However, when I checked my usage dashboard, I saw that this single request consumed a staggering 16 Million tokens! I have attached the screenshot for reference.
I am trying to figure out if there is something I am misunderstanding here. Did the Agent mode get stuck in a massive, unnecessary runaway loop?
I am currently a paying user (Ultra plan), but my quota for “Other Models” is running out in less than a week because of extreme anomalies like this.
What are the best practices to maintain a good “token economy” when using Agent mode so I don’t burn through my monthly limits so quickly?
Apologies if I am missing something obvious, but 16M tokens for a 3-line fix just seems unbelievable to me. Has anyone else experienced this, and how can I prevent it? Any help or clarification would be greatly appreciated.
Well, in the end, it’s not the number of changes made that determines the price, but the process leading up to it. So how long did the agent have to spend gathering information on their own, reading the codebase, possibly researching online… or simply guessing and making x attempts due to missing information in the prompt? The more (ideally useful) context you give the agent, the cheaper the fix will be in the end. The motto here is “garbage in, garbage out.” Too little information makes everything more complicated. But yes, tokens have generally become more expensive—not just in Cursor, but in other tools like Claude Code as well. That said, Anthropic’s models are definitely pricey, and Fable 5 is the most expensive of them all. The question is, would a smaller model have been sufficient in your case?
Well, that’s not normal. If what you did was just ask the model to fix this rule, it might have searched the whole codebase, exploring every corner and cranny. In open-source codebases, the context can be quite large, as there are often several different contributors who have made fixes and added code based on their own understanding of the codebase.
So, if you want to do this again with an open-source project, specify the exact line and provide the proper context and relevant files for the model to read. Ask it to understand those files, suggest a fix, and then implement it, rather than simply asking it to fix this. This way, you can avoid consuming a hefty amount of tokens on reasoning, context searches, and multiple iteration loops overall.
Also, if you already have a reasonable fix in mind, let the modal know. If the approach is viable, the model can use it as a starting point instead of exploring every possible solution, which can significantly reduce the amount of reasoning and make the overall process much more efficient.
Hey @Utku_Yilmaz
That 16M is mostly cache reads/writes and repeated automatic context compaction from a long-running chat, not fresh work on your fix. It shouldn’t pile up like that. We’ve let the team know and it’s an issue we’re tracking.
To avoid it: start a fresh chat per task, keep always-on rules/skills lean (they sit in context every turn), scope with @-mentions, and for a small fix point the agent at the exact file/lines. A lighter (non-Fable) model is usually plenty for quick edits. More here.
