Claude Code Burns 33K Tokens Before It Even Reads Your Prompt
Wire-level analysis shows Claude Code sends 33K tokens before reading your prompt - 4.7x OpenCode, with 3.7x higher real-task costs.
A developer placed a logging proxy between agentic coding CLIs and their model endpoints to measure raw token traffic instead of trusting built-in usage meters. The finding: Claude Code sends roughly 32,800 tokens before a single character of the user's prompt is even processed, while OpenCode sends only about 6,900 for the same task - a 4.7x gap. Most of that comes from system prompt size and tool schemas: Claude Code ships 27 tool definitions costing 99,778 characters, versus OpenCode's 10 tools at 20,856 characters.
The more damaging pattern shows up in caching. Across five requests for the same task, Claude Code generated 53,839 tokens of cache writes versus OpenCode's 1,003 - a 53.7x difference. The wire capture suggests Claude Code's cache breakpoint keeps shifting mid-session, forcing repeated full-price billing for context that should have been cached. On a ten-task benchmark where both harnesses passed every test, Claude Code used about 268,000 input tokens per passing run versus OpenCode's 72,000 - identical correctness, 3.7x the cost. Spinning up two subagents pushed Claude Code's cumulative tokens from 121K to 513K, a 4.2x multiplier stacking on top of the baseline overhead.
The researchers flag real limitations: single-machine testing, small sample sizes, and a snapshot that could shift as both tools evolve. Still, the actionable takeaways are clear: audit how many MCP servers you have wired in, watch instruction-file size, and treat subagent fan-out as a cost decision, not a free parallelization trick. If cost-per-task matters more than raw capability, this data is a solid reason to benchmark OpenCode against your real workload before assuming the flagship harness is the efficient default.