Claude Code token overhead dwarfs OpenCode
A comparative benchmark by Systima AI reveals Claude Code's baseline payload is ~33,000 tokens, nearly five times OpenCode's 7,000-token baseline. While Claude Code batches parallel tools efficiently on multi-step tasks, prefix instability triggers frequent cache re-writes, inflating costs up to 54x compared to OpenCode.
While Claude Code's batching is superior for complex multi-turn workflows, the sheer weight of its agent harness and its unstable caching strategy make it an expensive, context-hogging choice for simpler, single-turn tasks.
- –Claude Code's baseline of ~33k tokens consumes a sixth of a standard 200k context window before any codebase context or user prompt is read.
- –Caching is not a magic bullet: Claude Code's unstable prefixes cause frequent mid-session cache re-writes, which are billed at premium rates.
- –Multi-agent delegation is a major cost driver: fanning out to subagents creates a massive 4.2x multiplier on token usage because each subagent pays its own bootstrap cost.
- –Project configuration (MCP servers and instruction files) heavily inflates the baseline of both tools, with instructions adding around 20k tokens per request.
- –On multi-step tasks, Claude Code's ability to batch parallel tool calls helps it remain competitive with OpenCode by minimizing the total number of API turns.
DISCOVERED
1h ago
2026-07-12
PUBLISHED
3h ago
2026-07-12
RELEVANCE
AUTHOR
systima