Understanding Token Usage in Vibe Coding in 2025
By Mal Sherlock
I see lots of support questions about token usage. Though this might help explain it for some. I generally see 30k -500k tokens for changes in my apps.
Here’s a breakdown to help clarify why this happens and how you can optimize it.
Token Breakdown:
Input vs Output Input Tokens (The Heavy Hitter)
Most of your token usage comes from input tokens, which include:
-Full codebase: The AI needs your entire codebase as context.
-Change requests & specs: Detailed instructions and requirements.
-Conversation history: Previous chats and established coding patterns.
Output Tokens (Smaller Portion)
A much smaller portion of tokens is used for generating:
-New or modified code
-Explanations of changes
-Suggestions for testing or implementation
Why Selective Reading Isn’t Possible?
Its all about Interconnected Dependencies
-Components often rely on shared utilities and styles.
-Changes in one area can affect others.
-The AI must understand naming conventions, data flow, and state management.
Quality + Rework
-Comprehensive analysis is needed to avoid breaking your app.
-The AI checks for conflicts, ensures consistent patterns, and maintains type safety.
Growing Apps = More Tokens + Yes 500k is common
-Larger codebases require more context.
-Tightly coupled components and legacy code increase processing needs.
Why This Pattern Persists?
-AI starts fresh every time: Unlike humans, AI doesn’t retain memory between requests unless you provide the context.
-Safety-first approach: The AI prioritizes not breaking your app, so it processes more information to be thorough.
-Architecture limitations: Current models can’t perfectly isolate only the relevant context, so they include more than necessary to be safe.
Bottom line: High token usage is the trade-off for reliability and quality in current AI-powered coding.
Easysite is awesome. For me, $10–$100 covers 5M–50M tokens—enough for several apps. Still cheaper and more effective than many alternatives!
