Users discovered that when using Claude Code combined with Gemini-3-pro, a single conversation consumes up to 14.9k tokens, far exceeding expectations. This issue involves the AI Agent’s Context pre-reading mechanism, with users speculating that the project structure or code is pre-read at startup. They’re asking whether this high consumption is normal and how to reduce Context pre-reading volume to optimize resource usage. The article also compares the user experience between Claude and Cursor, shares stress tests on 2api, and seeks recommendations for MCP plugins. This discussion reveals performance challenges of AI tools in actual development, providing developers with practical optimization advice and offering technical depth and industry insights.
