Claude Code Sends 33k Tokens Before Reading Your Prompt: The Hidden Cost Heavy Users Must Know

If you run Claude Code against the API and watch your usage dashboard, you have probably noticed sessions costing more than expected. A new independent benchmark published this week gives the clearest explanation yet: Claude Code sends roughly 33,000 tokens of system prompt, tool schemas, and injected scaffolding before your actual prompt even arrives at the model.

For comparison, OpenCode, the open-source alternative, sends about 7,000 tokens. That is a 4.7x overhead gap on Claude Sonnet 5, measured at the API boundary with a logging proxy intercepting every request.

The finding comes from Systima AI, which spliced a proxy between each coding harness and the Anthropic endpoint, capturing exact JSON payloads and usage blocks on identical tasks.

Where the 33,000 Tokens Go

The breakdown is worth understanding because it changes how you should structure your workflow:

Tool schemas: 24,000 tokens. Claude Code ships 27 tools in its bootstrap. Beyond the coding core, it includes an orchestration suite covering background agents, CronCreate, Monitor, Task delegation, worktree management, and push notifications. OpenCode ships 10 tools and spends about 4,800 tokens on schemas. Claude Code also carries a system prompt of roughly 26,891 characters (6,500 tokens), versus OpenCode’s 8,811 characters (2,000 tokens).

CLAUDE.md adds another 20,000 tokens per request. A 72KB instruction file, typical for a production repo, adds roughly 20,000 tokens to every single API call. This is not a one-time setup cost: it is re-sent on every turn unless prompt caching absorbs it.

MCP servers compound the problem. Five modest MCP server definitions add 5,000 to 7,000 more tokens. Add everything together and a real working setup is 75,000 to 85,000 tokens deep before you type a word.

Subagent mode multiplies per-task cost by 4x or more. A task that cost 121,000 tokens run directly cost 513,000 tokens when fanned out to two subagents. Each subagent pays the full bootstrap cost independently, and the parent then consumes the full transcript.

The Cache Write Premium Nobody Talks About

A common assumption is that prompt caching makes the overhead free. It does not, for three reasons:

First, cache writes are billed at a premium over standard input tokens, not at a discount. You pay more on the first turn to write the cache, then read it back cheaply on subsequent turns. If a session is short, you pay the write premium without collecting enough reads to break even.

Second, Claude Code re-writes cache tokens mid-session, run after run. Systima found it wrote up to 54x more cache tokens than OpenCode across the same task set. That is 54x more cache write billing events.

Third, context-window consumption is unaffected by caching. A 33,000-token baseline means every turn starts one-sixth of the way into a 200,000-token window before any code enters the conversation. On a long session involving a large repo, you hit context limits faster and more frequently than a leaner baseline would.

When Claude Code Actually Wins on Cost

The full picture is more nuanced than “smaller baseline is always cheaper.” On the multi-step write-run-test-fix loop, Claude Code batched an entire job into a single parallel tool round trip. OpenCode serialized it across nine requests, re-paying its 7,000-token baseline nine times. The totals converged.

The formula is: total cost roughly equals baseline times request count, plus conversation growth. A large-baseline harness that batches aggressively can match a small-baseline harness that serializes. The meter starts higher with Claude Code; how the session unfolds decides who spends more.

For single-turn tasks and short bursts, OpenCode wins clearly. For complex, multi-step agentic workflows where Claude Code’s orchestration tools earn their keep, the advantage is less obvious.

The Model-Conditional Wrinkle

Systima’s finding that surprised them: on Claude Fable 5, the gap shrinks to 3.3x. Claude Code’s system prompt is model-conditional. It sent 27,787 characters of doctrine to Sonnet 5 but only 10,526 to Fable 5, with tool schemas also trimmed. OpenCode’s payload was byte-identical across both models.

This means the cost math changes depending on which model you route to inside Claude Code. If you are on Fable via the Max subscription, your baseline overhead is materially lower than the 33k Sonnet figure.

What This Means for Your API Bill

Let us run the numbers for a heavy user doing 200 Claude Code tasks per day at the API:

Scenario A: Simple file-read tasks, API pricing

• Claude Code: 199,000 tokens per task (measured) x 200 tasks = 39.8M tokens/day
• At Sonnet 5 intro pricing ($2/M input): $79.60/day, $2,388/month
• OpenCode: 41,000 tokens per task x 200 tasks = 8.2M tokens/day
• At Sonnet 5 intro pricing: $16.40/day, $492/month

Scenario B: Multi-step agentic tasks where Claude Code batches

• Both tools converge toward similar token counts per completed task
• Claude Code’s advantage: fewer HTTP requests, more reliable batching
• OpenCode’s advantage: predictable, auditable payload size

For users running simple to medium tasks, the overhead gap is real and significant. For users running complex agentic pipelines where Claude Code’s 27-tool orchestration suite earns its keep, the per-task cost often converges.

The critical variable is your mix. If you use Claude Code primarily for quick edits, file reads, and short sessions, you are paying a substantial infrastructure tax for orchestration features you are not using.

Practical Levers for Heavy API Users

Slim your CLAUDE.md. Every kilobyte you remove from your project instruction file saves roughly 250 tokens per API call. A 72KB file costs 20,000 tokens; trimming it to 20KB saves 13,000 tokens per request, every request, for the life of the session.

Audit your MCP server list. Each server definition adds 1,000 to 1,400 tokens to the bootstrap. Load only what the current task needs. Systima’s data suggests five servers add 5,000 to 7,000 tokens: remove servers you are not actively using.

Prefer Fable 5 routing over Sonnet 5 when both are available. Claude Code’s model-conditional prompt trimming means Fable incurs a 3.3x overhead versus Sonnet’s 4.7x. If Fable’s capability is sufficient for your task type, you cut your baseline cost by roughly 30%.

Use direct API calls for single-turn tasks. The overhead is only worth paying if you need multi-step tool orchestration. For “explain this function” or “generate a test for this class,” a direct API call with a tailored prompt will be materially cheaper.

Track subagent spawns. The 4x cost multiplier on subagent tasks means spawning agents has to deliver more than 4x the value to break even. Review whether your workflows actually require parallel delegation.

The Broader Pattern

Claude Code’s 33,000-token bootstrap is not a bug: it is the price of shipping a fully-featured agentic orchestration platform inside a coding tool. The 27 tools, background agents, push notifications, and worktree management are all real features that enterprise teams use.

The problem is that the pricing is not transparent at the point of use. When you run a quick task in Claude Code and your usage dashboard ticks up unexpectedly, you are paying for the full orchestration infrastructure regardless of whether you touched it.

For teams spending $300 or more per month on AI coding tools, understanding this baseline cost is essential. The question is not whether Claude Code is worth paying for overall, but whether the specific task mix you are running justifies the overhead premium versus a leaner alternative.

The Systima benchmark gives you the data to make that calculation. The decision depends entirely on which half of your workflow you care about: short-burst productivity where OpenCode wins, or complex agentic pipelines where Claude Code’s batching advantages offset the larger baseline.