Spotify cut Claude Code tokens 90% with a 350-line routing rule, not a model breakthrough
krishnan · x · 2026-09-08
krishnan breaks down how Spotify achieved 90% token savings in Claude Code via routing, not modeling: its Shunt plugin acts as a pre-tool hook that intercepts tool calls—untargeted reads over 350 lines are blocked and rerouted to a cheaper worker model via Portal, while targeted reads still reach Claude and predictable code generation writes straight to disk without consuming context.
The key insight: asking a model to route itself is a suggestion, but a pre-tool hook is an enforced economic policy—reserving expensive reasoning for debugging, architecture, and other judgment-heavy work.
Caveat: the reported 90% is mean bulk-read token savings across four scenarios in a Java monorepo, not a 90% cut in total engineering cost, and delegation adds 10–30 seconds of latency.
Related event: Spotify's Shunt plugin cuts Claude Code token usage by 90%(5 posts)→
More from coding & agent
- OpenAI rejected Meta-style internal tooling teams, betting on AGI-first — and Codex proved it right — dotey · 2026-09-08
- Anthropic's Claude Cookbook repo hits 52.5k stars: official notebooks for agents, skills and more — Roger_M_Taylor · 2026-09-08
- Routing LLMs across GitHub Copilot and Azure Foundry: same prices, but VS credits as backup — unixterminal · 2026-09-08
- Jeff Dean releases 1-hour AI engineering lecture, from LLMs from scratch to one human coordinating 100 agents — Roger_M_Taylor · 2026-09-08
- New coding-agent benchmark: Claude Opus 5 passes 23.9% vs 82.2% human expert — dair_ai · 2026-09-08
- Pro tip: enable experimental context management for gpt-6-astra in Codex, 400k tokens sweet spot — brandon_galang · 2026-09-08