Measured finding

31,428 tokens before a word of yours.

With the tools option omitted, this SDK/runtime sent all 27 built-in tool definitions. Including them increased the estimated input by 29,645 tokens, or 94.3% of the captured total.

What filled the request before inference

Source: captured SDK request and Anthropic's official count_tokens endpoint

  • Tools included 29,645 tokens · 94.3%
  • Skill descriptions 1,486 tokens · 4.7%
  • Base request 301 tokens · 1.0%

Scope

One isolated capture: SDK 0.3.168, runtime 2.1.168, model claude-opus-4-8, the tools and systemPrompt options omitted, settingSources: [], and no claude_code preset.

Reproduce

Every request, response, hash, package version, and measurement input is published for inspection.

Loaded context catalog

  1. Tool definitions

    The built-in tool objects supplied to the request.

    Items: 27
  2. Skill descriptions

    Description metadata loaded up front; full SKILL.md bodies load only when invoked.

    Items: 13
  3. Base request

    What remains after removing tools and skill-description text.

    Items: 6

Per-item token counts are independent comparison estimates. They rank items; they do not reconstruct the overview total.