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.