Anthropic's Claude Code sends roughly 33,000 tokens of pre-prompt context before a user types a single character, consuming 4.7 times more baseline payload than open-source alternative OpenCode. Research published by Systima.ai shows that heavy system prompts, tool schemas, and unoptimized prompt caching significantly increase developer usage costs. Understanding this hidden token overhead helps engineering teams manage API budgets and optimize context window efficiency.
Measuring the Fixed Baseline Gap
To isolate baseline payload overhead, Systima.ai tested Claude Code version 2.1.207 against OpenCode version 1.17.18 on identical tasks using a logging proxy between the harnesses and model endpoints. On a minimal task requesting a one-word response, OpenCode sent roughly 6,900 tokens, whereas Claude Code transmitted nearly 32,800 tokens on its initial turn.
The difference stems primarily from tool definitions and behavioural scaffolding. Claude Code ships with 27 tools—including background orchestration, cron tasks, and worktree management—accounting for nearly 24,000 tokens. In contrast, OpenCode defaults to 10 classic coding tools consuming approximately 4,800 tokens. Even when all tools were disabled, Claude Code's system instructions remained over three times larger due to strict behavioural doctrine and safety rules.
How Configuration Files and Subagents Scale Overhead
In real-world repositories, baseline token consumption expands rapidly. Systima.ai found that dropping a 72KB repository instruction file (CLAUDE.md or AGENTS.md) added an average of 20,000 tokens to every single API request. Adding five Model Context Protocol (MCP) servers appended another 5,000 to 7,000 tokens, pushing initial payloads in configured environments to between 75,000 and 85,000 tokens before user input.
Subagent orchestration creates an even larger multiplier effect. A multi-step task costing 121,000 tokens directly surged to 513,000 tokens when fanned out to two subagents. Because each subagent operates as an isolated instance, it re-reads its entire baseline and tool schema on every turn.
Key Benchmark Findings
- First-turn overhead: Claude Code requires ~32,800 pre-prompt tokens compared to OpenCode's ~6,900 tokens.
- Cache rewrites: Claude Code re-wrote mid-session prompt-cache tokens frequently, generating up to 54x more cache write tokens than OpenCode on identical tasks.
- Instruction file impact: A 72KB instruction file adds roughly 20,000 tokens per request across both harnesses.
- MCP server tax: Attaching five public MCP servers increased tool schema weight by 4,900 to 6,967 tokens.