Does anyone else feel like whenever you run into an error and ask for a fix—or explain the how, where, what, and why—even if you’re super specific, resolving it somehow always blows up into costing over 20k tokens? Even if it’s just deleting a newline in a 20-line or 500-line HTML file, and in doing so it rewrites the issue, basically regenerating it or bringing it right back. I’m sick of it. I’ve already tried countless ways to develop with AI, but it’s always the same thing. No matter the prompt tweaks, SDDs, skills, rules, limiters, agents, and all that BS they talk about, you always get the same basic, generic results and errors—super flat, plain, completely homogeneous. They release a new model and it just does the exact same thing a level higher; like, if it used to mess up a div, now it messes up a switch case (just as a funny example). I don’t know, it’s always the same deal: driving costs through the roof—meaning your own expenses in tokens, time, and the energy fed to them and to training their products. At least put some skin in the game and give us some breathing room so it’s a bit fairer and more balanced, and we can actually squeeze out slightly better profit margins before we get replaced by these half-baked tools.
Make sure you aren’t overloading the context window with unnecessary skills, plugins, etc - more is not better. You might want to try from scratch, nothing else enabled, and compare. The newer versions of the Antigravity interface will give you counts of the tokens used per skill/plugin/etc, so you can help narrow down where problem areas might be as well.
Hi @Mauro_Oscar,
A few things that usually cut the waste a lot on small fixes:
- Start a fresh conversation for small fixes. Long chats carry a lot of history, so every new request costs more and the agent is more likely to pick up old mistakes again.
- Point it at the exact spot. @mention the file and give the line, e.g. “In index.html, line 42, remove the extra line break. Don’t change anything else.”
- Ask for a plan first. Try “Tell me what you’ll change before editing” so you can stop it before it rewrites half the file.
- Use a lighter model for quick edits. Flash is faster and cheaper for small changes. Save the bigger models for real problem-solving.
- Do tiny edits yourself. For something like a single line break, editing by hand is honestly quicker.
And when it gets something wrong, please hit the thumbs-down on that response. That feedback goes straight to the team and helps improve exactly these cases.