Agent router log: 8,458 escalations up, zero down — caching beat routing by 70%
mrtrly · reddit · 2026-09-22
The author built a model router to cut agent costs, then audited its behavior: across 59,299 requests (list price $5,085, median 4 cents), 8,458 escalated up and zero went down — potentially 30% of spend recoverable.
What actually saved money was caching: uncached list price would have been $16,980 vs $5,085 cached (70% off), and escalation breaks the cache at roughly 7x cost. Worst of all, 72% of coder spend ($3,218) went to canceled tasks; one "fix broken tests" task ran 76 attempts, burning $940, with the deepest autofix chain reaching 19 levels.
His takeaway: when your default model is already cheap, a router is a quality dial, not a savings dial — cost per shipped item matters, cost per request is vanity. The logging proxy is open source.
More from coding & agent
- User Vibecodes Tampermonkey Scripts to Patch Long-Ignored UX Bugs in University Systems — craig_macdonald · 2026-09-22
- Speaking UI layouts into existence with Jev, shadcn and a local transcription model — _AustinCalvert_ · 2026-09-22
- ai-copywriter hits 1k stars: an agent skill that writes converting copy with zero AI tells — tom_doerr · 2026-09-22
- ContextBridge: an open-source pool that routes AI tasks across local machines, APIs and shared hosting — IamAngusU · 2026-09-22
- Microsoft's 44-page playbook: 100+ internal AI projects show licenses alone fail — alex_verem · 2026-09-22
- signal-mcp: open-source local-first MCP server exposing Signal to AI clients — googlarz · 2026-09-22