Claude Code sends 33k tokens before reading the prompt

Claude Code has been sending 33,000 tokens before even reading user prompts, according to new research from [Systima AI](https://systima.ai/blog/claude-code-vs-opencode-token-overhead). This massive overhead occurs before any actual work begins, creating significant cost implications for developers building AI applications.

The behavior represents a fundamental inefficiency in how some AI coding assistants operate. While users expect to pay for the AI's output and processing, they're unknowingly paying for extensive background operations that happen before their prompt gets processed.

Why Token Overhead Matters

Tokens are the basic unit of measurement for AI model usage, and each token costs money. When an AI system burns through 33,000 tokens before starting work, it's like a contractor charging for extensive prep work that the client never requested or approved.

This overhead becomes especially problematic with sub-agents - specialized AI components that Claude Code spawns to handle different aspects of complex tasks. Users report situations where Claude Code immediately launches seven sub-agents for a single request, each carrying the same token overhead. The result is budget depletion before any meaningful work gets completed.

The issue extends beyond large system prompts. Modern coding agents increasingly trigger excessive tool usage even for simple requests. Some systems now execute 30+ tool calls for basic prompts like "Hey" or "commit," multiplying costs for trivial interactions.

The Transparency Problem

Token overhead highlights a broader transparency issue in AI coding tools. Users often can't see what's happening behind the scenes or why their credits disappear so quickly. This opacity makes it difficult to optimize usage or understand true costs.

Different AI coding assistants handle this differently. Some systems allow users to inspect the full agent system prompt and understand exactly what they're paying for. Others operate as black boxes, leaving developers to discover inefficiencies through trial and expensive error.

Caching offers one potential solution. After the first interaction, those 33,000 tokens should become cache hits, reducing costs to one-tenth the original price. However, this benefit only applies after the initial expensive interaction, and not all systems implement caching effectively.

Impact on Development Economics

This inefficiency pressures developers to carefully evaluate AI coding tools based on total cost of ownership rather than just advertised per-token pricing. A tool with higher per-token costs but lower overhead might prove more economical than one that appears cheaper but burns tokens on background operations.

The overhead problem also breaks the cost-effectiveness equation for many AI applications, particularly those involving frequent, small interactions. When simple requests trigger massive token usage, the economics of AI-assisted development shift dramatically, potentially making human developers more cost-effective for certain tasks.