aiexpert
Home / News / Brief
Breaking · Aug 18, 2026, 07:35 PM · 2 sources

Anthropic cuts Claude Code built-in API skill from 200k to 25k tokens via on-demand loading

Anthropic reduced the context cost of Claude Code's built-in claude-api skill from ~200,000+ tokens to ~25,000 tokens by loading reference documentation on demand instead of including it in every session context. The optimization is part of a broader series of August 2026 updates to Claude Code that improve mid-turn command support, error messaging, and transcript rendering.

The claude-api skill is invoked whenever Claude Code needs to access Anthropic's API reference, which happens frequently when building agents or integrating Claude into applications. At 200k-token overhead per session, it represented a meaningful burn rate for subscription users (who have limited weekly token budgets) and API users running long coding sessions. The on-demand loading pattern trades a small lookup latency for 87.5% context reduction, making it roughly equivalent to a fixed $2-3 savings per session at Sonnet pricing.

For Claude Code users: this reduction directly improves session economics, especially for developers hitting weekly usage limits. Combined with /compact (mid-session summarization) and prompt caching, the improvement highlights how context management is becoming the primary lever for token efficiency in agentic coding workflows, not model capability.

Sources

Everything this brief rests on
  1. 01 Primary source releasebot.io
  2. 02 releasebot.io releasebot.io “Reduced context cost of loading built-in claude-api skill from ~200k+ tokens to ~25k by loading reference docs on demand”