DevOps / SRE / Platform · 18.08.2026, 20:40 UTC
A Claude Code skill was eating 200,000 tokens before answering a single question
| Schweregrad | info |
|---|---|
| Kategorie | DevOps / SRE / Platform |
| Quelle | The New Stack ↗ |
| Veröffentlicht | 18.08.2026 UTC |
Sicherheitsmeldung mit Schweregrad noch nicht bewertet. Technische Details im Tab „Originaltext“; empfohlene Schritte in der Checkliste.
A Claude Code skill designed to help developers work with Anthropic’s API was consuming more than 200,000 tokens to load. With Claude Code v2.1.234, Anthropic says it has brought that down to roughly 25,000 tokens.
The change appeared in the Claude Code changelog on Monday. Anthropic attributed the reduction, which cuts the initial context cost by at least 85.7%, to loading the skill’s reference documentation on demand. In this case, a single bundled skill could consume more tokens than many coding sessions do from beginning to end.
Skill loaded everything upfront
The built-in /claude-api skill which loads reference material for developers working with the Claude API and Managed Agents, can be invoked directly, but Claude Code can also activate it when a project imports Anthropic’s Python or TypeScript SDK. Interestingly, developers had already traced the problem before Anthropic documented the fix.
In a GitHub issue opened July 7, a developer examining Claude Code 2.1.201 found that /claude-api embedded its shared reference files and the detected language documentation directly into the skill body. The report measured roughly 120,000 tokens of reference material in a single invocation. One migration document accounted for an estimated 36,000 tokens on its own.
One migration document accounted for an estimated 36,000 tokens on its own.
Once the rest of the skill loaded, even a one-line question could consume roughly 200,000 tokens before Claude started answering — the hidden overhead that, as with production AI pipelines more broadly, only surfaces once someone …