{"rewrite":{"id":"r_a945d4a414fd7d93b822008c","clusterId":"c_e6e8bc170cb869df72ce505f","slug":"claude-code-uses-33-000-tokens-before-user-input-and-75-000-in-full-setup","model":"deepseek-v4-flash","headline":"Claude Code Uses 33,000 Tokens Before User Input and 75,000 in Full Setup","summary":"AI consulting firm Systima measured Claude Code's token consumption and found it sends about 33,000 tokens before the user's first prompt, compared to about 6,900 for OpenCode. In a practical configuration with instructions and MCP servers, Claude Code's initial load reached around 75,000 tokens.","whyItMatters":"The high token overhead of Claude Code before user input raises operational costs and reduces the effective context window for developers using the tool.","webCardHtml":"\u003cp\u003eThe breakdown of Claude Code\u0026#39;s initial request shows tool descriptions alone account for about 24,000 tokens, while the system prompt is about 6,500 tokens. OpenCode, by contrast, uses roughly 2,000 tokens for its system prompt and 10 tool descriptions. Adding a 72KB instruction file like CLAUDE.md adds about 20,000 tokens per request for both tools. Each Model Context Protocol server adds around 1,000 to 1,400 tokens. In a practical configuration with instruction files, five MCP servers, and plugins, Claude Code sent about 75,000 tokens before the user input, with 118 tools in the transmitted data totalling 311KB. The test used Claude Sonnet 4.5 as the AI model.\u003c/p\u003e","blueskyPost":"Claude Code sends about 33,000 tokens before a user's first prompt, rising to 75,000 in a full configuration. OpenCode uses roughly 6,900. Systima's measurement shows tool descriptions account for the bulk.","twitterPost":"Claude Code sends about 33,000 tokens before user input, 75,000 in full config. OpenCode uses ~6,900. Tool descriptions account for 24,000 tokens. Systima's measurement.","threadsPost":null,"newsletterBlurb":"A Systima test measured Claude Code's token overhead. The tool sends about 33,000 tokens before the first user prompt, with tool descriptions taking up 24,000 of those. In a practical setup with instructions and MCP servers, the initial load reaches 75,000 tokens, compared to 6,900 for OpenCode.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260713-claude-code-consuming-token/\",\"title\":\"Claude Code Uses Approximately 33,000 Tokens Just for Initial Data Before Instructions, Reaching About 75,000 Tokens in Practical Configuration\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4223,"outputTokens":1872,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1783930768,"createdAt":"2026-07-13T08:09:16.000Z","publishedAt":"2026-07-13T08:12:28.000Z","updatedAt":"2026-07-13T08:09:16.000Z"},"cluster":{"id":"c_e6e8bc170cb869df72ce505f","canonicalTitle":"Claude Codeは指示前の初期データだけで約3万3000トークンを使用、実用構成では約7万5000トークンに","representativeArticleId":"a_36e84c218f1b4cf8849a38ed","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-07-13T07:20:00.000Z","lastSeenAt":"2026-07-13T07:20:00.000Z","updatedAt":"2026-07-13T08:12:28.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260713-claude-code-consuming-token/","title":"Claude Codeは指示前の初期データだけで約3万3000トークンを使用、実用構成では約7万5000トークンに"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":[],"studios":[],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":null}