Systima Measured Claude Code Using 4.7x More Tokens Than OpenCode Per Request — That Measurable Gap Is a Sellable Coding Agent Cost-Audit Service.

Systima did something most engineering teams only do internally when picking a coding agent: it measured the token overhead of two competing harnesses side by side, then published the exact numbers.

That is a gift to anyone who wants to sell AI agent cost audits.

What Systima actually measured

On a first request with tools loaded, "Claude Code used roughly 33,000 tokens of system prompt, tool schemas, and injected scaffolding before the prompt even arrived. OpenCode used about 7,000." Broken down: Claude Code's system prompt and 27 tool schemas ran to roughly 32,800 tokens (27,344 characters of system prompt plus 99,778 characters of tool schemas), against OpenCode's roughly 6,900 tokens (9,324 characters of system prompt plus 20,856 characters of tool schemas across 10 tools) — a 4.7x gap before either harness reads a single word of the user's request.

The gap compounds under real use. A production repository's 72KB instruction file (an AGENTS.md or CLAUDE.md) adds "another (avg) 20,000 tokens to every single request" in both harnesses. On a file-summarization run, Systima measured Claude Code writing 53,839 cache tokens against OpenCode's 1,003 — over 50x more, on tasks where both harnesses hit the same 5/5 pass rate on Systima's 10-lane test suite. The most expensive multiplier of all: "a small task that cost 121,000 tokens done directly cost 513,000 tokens when fanned out to two subagents."

Not every number favors OpenCode, either — on Systima's multi-step task, Claude Code's habit of batching tool calls into fewer requests brought its cumulative total to roughly 121,000 tokens across 3 requests, versus OpenCode's roughly 132,000 tokens across 9 requests. The full picture needs both numbers, not just the flashy ratio.

The business idea

Most teams running coding agents in CI or production never measure any of this. They pick a harness, wire it into their workflow, and keep paying the bill without knowing whether their specific setup — their instruction file size, their MCP server count, their subagent fan-out pattern — is burning tokens for no quality gain.

That gap is the service: a fixed-scope "coding agent token audit" that benchmarks a client's actual repo and workflow across the harnesses they're considering, using Systima's own methodology as the credibility anchor.

Money play

1. Sell a benchmark audit that measures a client's real instruction file, MCP server count, and subagent pattern against at least two harnesses, using Systima's finding that identical instruction files add roughly the same 20,000 tokens per request regardless of harness as your baseline control.

2. Lead with the subagent fan-out number as the highest-value catch: teams that don't know a two-subagent fan-out can turn 121,000 tokens into 513,000 are the easiest audits to sell, because the fix — route more work through direct calls, cap fan-out depth — is cheap once it's visible.

3. Offer a cache-efficiency pass as an add-on: Systima's 50x-plus cache-write gap on a file-summarization task shows that quality parity (both harnesses passed 5/5) doesn't mean cost parity — that's a billable finding, not a one-time observation.

4. Package a monthly re-benchmark retainer. Harness system prompts and tool schemas change with every release; a config that was efficient last quarter may not be today.

5. Use "every token of harness payload is a token of working context you cannot spend on your task" as the one-line pitch that gets a technical buyer to book the first call.

Bottom line

Systima didn't just settle a Claude Code vs. OpenCode argument. It published a reusable measurement method — system prompt tokens, tool schema tokens, cache writes, fan-out multiplier — that anyone can repackage as a paid audit for teams that have never once measured what their coding agent costs before it reads their prompt.

Sources:

https://systima.ai/blog/claude-code-vs-opencode-token-overhead

https://news.ycombinator.com/item?id=48883275