How much does an AI coding assistant cost before it does any actual work? According to a study published on the systima.ai blog, the answer depends heavily on which tool you pick.
The authors measured the "token overhead" of two AI coding assistants — the baseline volume of text each one sends to the underlying language model before it has even processed a user's request. They report that Claude Code sends roughly 33,000 tokens up front, while OpenCode sends about 7,000.
Tokens are the units of text that AI models read and generate, and they are also the units that get billed. A higher baseline means every single interaction starts from a more expensive floor, regardless of how simple the task is.
The systima.ai team says the investigation began as a hunch. They normally use OpenCode, but were "forced" to switch to Claude Code for a stretch due to issues with a tool they call Meridian. During that period, they say, they watched their usage meter climb much faster than it did on OpenCode. That anecdote prompted the small study that produced the token figures.
The post frames this as a difference in how much context each tool loads by default — system instructions, tool definitions, and other scaffolding — before the user's prompt is even considered.
Why it matters: as developers lean more heavily on AI assistants that charge by the token, a several-fold difference in baseline overhead can quietly translate into much larger bills for the same amount of work.