Anthropic Cut One Skill's Context Cost by 88% Without Deleting a Word

Buried in the Claude Code 2.1.234 changelog is a line worth more than most feature announcements: the built-in claude-api skill now costs ~25k tokens to load instead of ~200k+ — an 88% cut, achieved by loading its reference docs on demand rather than up front. Nothing was removed. The docs are all still there.

Sit with the number that was there before. 200k tokens is the entire context window on most models, and the whole budget on a 200k plan — spent by one skill, before you typed anything. The reference material wasn't wrong to exist. It was wrong to be resident.

In-house agents fail the same way, and almost nobody measures it. The system prompt carries the full API reference, the database schema, the style guide, and every tool definition, loaded on every turn because deciding what to load was harder than loading everything. That bill comes due on every single request. It also hides: nothing errors, latency drifts up, the model's attention thins across material it didn't need, answers get vaguer — and the team concludes the model got worse.

The fix is unglamorous. Ship an index, fetch the section when the task actually calls for it. Progressive disclosure isn't an optimization you do later; it's the difference between an agent that scales to real work and one that spends its window on documentation it never reads.

See how we close it →