Spotify exec shunts Claude's grunt work to cheap bot, slashes token bill 90%
Spotify product manager Dimitri Mazmanov describes a Claude Code plugin called shunt, which sends file-reading and boilerplate-writing tasks to a cheaper model, Gemini 2.5 Flash. It runs on two no-code AiKA modes in Portal by Spotify, plus hooks that block reads of files over 350 lines by default. In tests on a Java monorepo, bulk-read savings averaged about 90%. He says it can't handle editing or reasoning, such as spotting a thread-safety bug, and each delegation adds 10 to 30 seconds of latency.
Why it matters: It shows developers using AI coding tools may cut costs by routing simple reading and boilerplate tasks to cheaper models, though the approach adds delays and cannot handle editing or reasoning.
Comments
No comments yet.