[DigitalToday reporter Jinju Hong (홍진주)] Anthropic's AI coding tool Claude Code was found to send tens of thousands of tokens to an AI model even before processing a user's instructions. In real development environments, initial transmissions can reach about 75,000 tokens, an analysis said, potentially affecting not only costs but also use of the context window.

On July 13 local time, online outlet Gigazine reported that AI consulting firm Systima disclosed the result in a test comparing Claude Code and the open-source AI coding tool OpenCode.

The measurement ran Claude Code 2.1.207 and OpenCode 1.17.18 on the same computer environment using the same AI model, Claude Sonnet 4.5. After recording the data delivered to the model on a relay server, it calculated token counts.

The gap was large even in the simplest setup. In an empty working environment excluding both an MCP (Model Context Protocol) server and user settings, Claude Code sent about 32,800 tokens to the model in the first request when given the simple command, "Print OK only." OpenCode sent about 6,900 tokens.

The biggest difference came from the amount of tool descriptions available to the AI. Claude Code delivered descriptions for 27 tools to the model along with the system prompt, and the descriptions alone accounted for about 24,000 tokens. OpenCode included 10 tool descriptions, keeping total input at about 6,900 tokens.

The difference persisted even with all tool functions disabled. According to Systima, Claude Code's system prompt was about 6,500 tokens, while OpenCode's was about 2,000 tokens.

Such large initial inputs are not only an issue of higher costs. Based on Claude Sonnet's 200,000-token context window, Claude Code uses about one-sixth of the total context before reading actual code or user dialogue. Using a prompt cache can cut retransmission costs, but it does not reduce the tokens occupying the context window itself, Systima explained.

In real development environments, instruction files containing project rules were also found to add to the burden. Adding a 72KB 'CLAUDE.md' file increased both Claude Code and OpenCode by about 20,000 tokens per request. Because such instruction files are sent to the model with every request by default, token use rises in proportion as the file gets longer.

MCP servers that connect external services and development tools also increased the initial load. Because MCP must describe to the model the functions and arguments the AI can use, each additional server increased tokens by about 1,000 to 1,400. Connecting five MCP servers raised initial input by about 5,000 to 7,000 tokens.

Systima measured a real development environment including MCP servers, instruction files and plugins and found Claude Code sent about 75,000 tokens to the model before the user even entered a command. In that environment, the tools available to the AI rose to 118, and the data sent to the model was 311KB. Depending on the configuration, the first request can increase to about 75,000 to 85,000 tokens, Systima analyzed.

Even so, a high initial token count did not necessarily mean it was inefficient over the whole task. In a test that repeated file creation and program execution, Claude Code bundled and called multiple tools at once, cutting the total number of requests to three. Total input tokens were about 121,000 in that case.

OpenCode, by contrast, called tools one by one, sending a total of nine requests. Input tokens were about 132,000, which was higher. This means that even if the initial load is small, a higher number of requests may mean there is not a large difference in final token use.

By contrast, costs rose sharply when using sub-agents. When Claude Code split a task into two sub-agents, each agent reloaded the system prompt and tool descriptions, and an additional process in which the top-level agent rechecked the results pushed total input tokens up from about 121,000 to 513,000.

Systima said the results were measured in a specific environment and could vary depending on model versions and MCP configurations. It advised that when operating AI agents, users should not just check the pricing screen but record the prompts and token use actually delivered to the model, and add instruction files and MCP servers one by one to check the initial load.