What's inside Claude Code's startup token cost
Tokens sent at startup (first request, before any prompt is processed)
Systima measurements (July 2026; Claude Code 2.1.207 / OpenCode 1.17.18; model: Claude Sonnet 4.5).
Claude Code is a terminal-based AI coding agent from Anthropic. For this study, Systima ran Claude Code and OpenCode on the same machine and the same model, capturing the full payload actually sent to the model through a relay server to compare each tool's "invisible" consumption.
About 33,000 tokens before the prompt is processed
The task used for measurement was a mere 22 characters: reply with exactly "OK." Even so, Claude Code sent about 33,000 tokens — system prompt, tool definitions, and auto-injected reminders — before that instruction was ever processed. OpenCode sent about 7,000, a roughly 4.7x difference.
"When we asked both harnesses for a one-line reply, 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." — from Part I. The floor
At 33,000 tokens, that is roughly one-sixth of a 200,000-token context window (the amount of information the model can handle at once). Part of the window is already filled before any code or conversation is loaded.
Most of it is the 27 tool definitions
So what takes up that much? The largest term is the tool definitions. About 24,000 of the ~33,000 tokens were the descriptions of Claude Code's 27 tools — things like file editing and search. OpenCode has 10 tools at about 4,800 tokens, which is the main source of the gap.
"Roughly 24,000 of Claude Code's ~33,000 tokens are tool definitions, versus roughly 4,800 of OpenCode's ~6,900." — from Part I. The floor
Even with every tool stripped out, the system prompt alone was about 6,500 tokens for Claude Code versus about 2,000 for OpenCode. Claude Code carries more behavioral guidance — tone rules, task-management conventions — which makes it heavier.
A real setup balloons to the 75,000-token range
What gets added on top in a real setup (Systima measurements)
The measurements weren't limited to a clean environment; they also covered a setup close to real development work. What matters here is the per-project instruction file and MCP servers (a mechanism that connects external functions to the AI).
What the instruction file and MCP servers add
A 72KB instruction file added about 20,000 tokens on average, and five MCP servers added 5,000 to 7,000 — on every single request. As a result, the first request in a real setup reaches 75,000 to 85,000 tokens. In Claude Code's test configuration (four MCP servers plus plugins and an instruction file), the tool count swelled to 118 and the payload to 311KB.
"A production repository's 72KB instruction (AGENTS.md or CLAUDE.md) file adds another (avg) 20,000 tokens to every single request. Five modest MCP servers add 5,000 to 7,000 more. By the time a real working setup sends its first request, it is 75,000 to 85,000 tokens deep before the user has typed a word." — from Part II. The multipliers
In other words, the tool sets the floor, but your configuration sets the bill. With the same Claude Code, the amount sent per request can more than double depending on your settings.
Splitting work across subagents multiplied total usage ~4x
Another striking result is the measurement when work is split across subagents (a mechanism where a parent AI delegates to child AIs). A task that cost 121,000 tokens run directly ballooned to 513,000 — about 4.2x — when fanned out to two subagents. Each child agent re-reads its own system prompt and tool definitions on every turn.
"A small task that cost 121,000 tokens done directly cost 513,000 tokens when fanned out to two subagents." — from Part II. The multipliers
Startup cost isn't total cost — how to read the numbers
Cumulative token usage on a real (multi-step) task
Systima measurements. The heavier-at-startup Claude Code came in below OpenCode on total usage.
Looking at these numbers alone, it's tempting to conclude Claude Code is the expensive tool. But the measurements also show the opposite. The numbers need careful reading.
On a real task, the totals reversed
Because the baseline is re-sent on every request, total usage is roughly baseline times request count. On a multi-step task, Claude Code — which finished in three requests by batching operations — used about 121,000 tokens, while OpenCode, split across nine, used about 132,000. On the cumulative total, they reversed. A tool that's light at startup can still run up the bill if it takes many small round trips.
"OpenCode paid its ~7k baseline nine times, Claude Code paid its ~33k three times, and the totals converged." — from the article
Re-running with the model switched to Claude Fable 5, the startup gap narrowed from about 4.7x to about 3.3x. Claude Code varies how much guidance it sends by model, so the ratio isn't fixed.
"The floor gap on Fable comes out at roughly 3.3x by payload against 4.7x on Sonnet. Still far hungrier, but the ratio is model-dependent." — from the article
To cut usage, start by auditing your configuration
A practical note follows. The takeaway from this study is to review your own configuration before switching tools. A bloated instruction file and too many MCP servers keep piling onto every single request. Dropping unused MCP servers and keeping the instruction file lean can cut tens of thousands of tokens per request.
Small behavioral quirks matter too. At the time of testing, Claude Code (version 2.1.207) didn't load an instruction file named AGENTS.md and only ingested it when named CLAUDE.md. Whether your carefully written instructions are being ignored over a filename is worth checking. For Claude's per-model pricing and how to pick between them, see our Claude model cost comparison, and for getting started with Claude overall, see our Claude AI guide.
"Claude Code 2.1.207 ignored AGENTS.md entirely and only ingested the file when renamed CLAUDE.md, injecting it into the first user message." — from Part II. The multipliers
Conclusion: the heavy startup is real, but total cost is up to how you use it
Systima's measurements confirm, with numbers, that Claude Code sends about 33,000 tokens before processing a prompt and reaches the 75,000-token range in a real setup. At the same time, on an actual task the smaller request count let it come in below OpenCode — so the size of the startup cost alone doesn't settle which tool is cheaper. The thing to judge by is what your own configuration actually sends. Auditing your instruction file and MCP servers is the surest first step to cutting the bill.
