33k vs 7k tokens: what the comparative overhead reveals about Claude Code and OpenCode

In plain terms

A Systima.ai benchmark (July 12, 2026) measured what Claude Code and OpenCode send before processing the user prompt: ~33,000 tokens for Claude Code, ~7,000 for OpenCode. The difference is not cosmetic—it is architectural.

The fact

Systima intercepted the API between the CLI and the model: Claude Code sends a system prompt + a tool description + a usage reminder totaling ~33k tokens; OpenCode sends ~7k tokens for the same setup. Over an average usage cycle, this translates into additional billed tokens and pressure on the KV cache.

Analysis

Two philosophies oppose each other, and they are not neutral.

Claude Code plays the heavy harness: detailed instructions for tooling, embedded examples, rich role context, explicit safeguards. The bet: a long system prompt reduces behavioral variance, improves rule compliance, saves iterations. Cost: input tokens boosted with each turn.

OpenCode plays the light harness: minimal tooling, short system, let the model infer. The bet: fewer pre-fabricated anchors, more flexibility, lower bill. Cost: more variance, more risk of behavioral deviation when the task deviates from the happy path.

Under the hood

On pricing: at ~$3/M input tokens for Claude, 33k tokens per turn × N conversation turns, that's several cents per session that the user never sees. On the KV cache: Claude Code uses Anthropic's prompt caching—the prefix is cached and billed ~10% after the first time. That's the real economic lever. OpenCode, being shorter, doesn't cache much.

The "fair" comparison is not raw tokens but effective billed tokens (with cache) × output quality × number of retries. Systima measures only one of the three. The actual billing difference is probably closer to 2× than 5×, but remains significant.

Scenarios

• Convergence (50%): Claude Code lightens its preamble, OpenCode adds examples; both converge to 10-15k tokens.
• Assumed Divergence (35%): Claude Code remains positioned as "heavy harness, high obedience" (enterprise), OpenCode remains "light harness, user control" (independent).
• New Pattern (15%): Emergence of a modular harness (conditional loading of tools based on the prompt).

So what

The harness becomes a product differentiator. Comparing two agent CLIs is comparing two theses on what the model should know before reading the user. For a builder: measure your own preamble and its effective cost with cache. For a leader: perceived latency and monthly billing depend on an architectural choice that no one mentions in the commercial demo.

To watch

The publication by Anthropic or OpenAI of a "harness reference"; the emergence of tools that measure the real overhead per CLI; the appearance of official "diet" modes for Claude Code or Cursor.