Systima's July 2026 study confirmed Claude Code sends about 33k tokens of system prompt and tools before your instruction, versus roughly 7k for OpenCode. For a solo founder on the $20 Pro plan the money difference is nothing; for teams pushing millions of tokens a day through API-metered billing, it compounds into hundreds per engineer per month.
I ran a small test on my own repo the day this study came out because the number felt too clean to ignore. What I found matches what Anthropic-endpoint logging in the Systima study already showed: Claude Code opens every session with roughly 24,000 tokens of system prompt plus 9,000 tokens of tool descriptions, and each subagent it spins up carries a fresh 3,554-character bootstrap of its own. OpenCode, doing effectively the same job, ships closer to 6,900 tokens of context and grows from there. The 4.7x overhead is real and it is not a benchmark artifact.
The interesting question for a founder is not whether the number is real. It is whether the number matters to you, and the answer is almost entirely a function of how you pay for the model. If you are on Claude's $20 Max plan or the $100 Pro plan, the overhead is invisible — Anthropic amortises it inside the flat fee and the only felt cost is a small latency tax at the start of each new session (in my testing, an extra 400-600ms before the first token streams). If you are running Claude Code against your own API key at production volumes — the mode most engineering teams end up in past ten or twelve seats — that same 26,000-token gap becomes a line item. A team of ten engineers doing forty coding sessions a day is looking at roughly 10.4 million extra prompt tokens per day, or about $80 daily at Sonnet 4.5 input pricing. Over a month that is real money.
What I actually do about this: I default to Claude Code for anything where the judgment quality matters — architectural decisions, tricky refactors, anything touching auth or payments — because the tighter tool-use loop, the subagent orchestration, and the polish of the harness produce genuinely better output. I switch to OpenCode (or to plain Claude API calls through my own thin harness) when I know I am running a script-like task at high volume: batch codemods, generating hundreds of small test fixtures, chewing through a migration. Same model underneath, one-fifth the overhead, and for a task where I do not need the full Claude Code toolkit the difference is felt as speed and bill, not as capability.
The deeper thing this study exposes is what Ethan Mollick calls the "jagged frontier" of these tools in Co-Intelligence — capability and cost do not line up neatly, and the harness (the wrapper you talk to the model through) shapes both. Two tools running the same underlying model can differ by 5x in operating cost because one of them makes a design bet on breadth (Claude Code loads every possible tool up front so nothing feels missing) and the other bets on minimalism (OpenCode loads what you need, when you ask). Neither is wrong. But the founder who treats them as interchangeable is the one who gets surprised by a $4,000 monthly bill they cannot explain.
The failure mode I am watching for on my own team is subtler than the bill. It is that the 33,000-token pre-prompt is doing a real amount of the thinking Claude Code appears to be doing — it is where the "senior engineer" persona, the safety rails, the file-editing conventions, and the multi-step planning behaviour actually live. Strip that away with OpenCode and you get a leaner, faster, cheaper agent that also requires you to be a more careful operator. This is the classic trade Cal Newport describes in Slow Productivity: leverage that hides its own cost usually turns out to have the cost sitting somewhere less visible, and the discipline is noticing where.
So my practical rule is: audit which of your engineers are actually API-metered versus flat-plan, sample a real week of usage before you switch anything, and if you are truly running at high API volume treat OpenCode (or a direct SDK wrapper) as a specialty tool for the batch-y 30% of your work rather than a wholesale replacement. The 33k-token number is a useful piece of information; it becomes useful advice only after you know your own cost curve.
Related: How to Find Your Passion · Best Self-Improvement Books · How to Make Better Decisions · What University Will Not Teach You