Routing to smaller models may cost more: harness overhead adds 30-89%
yb2698 · x · 2026-10-23
The author tested routing easy queries from Qwen3.8-27B to Qwen3.5-9B via a near-free decision model (d3) to cut costs, and shared per-trial cost results across harnesses.
- IPython agent: +30% cost vs 27B alone
- Pi: +89%
- Claude Code: -1%, but with a 17k-token prompt
- Qwen Code: -7%, with a 34k-token prompt
Follow-up runs sampled 10 hard tasks from TB-Lite (3 attempts each), comparing harnesses on the large model against pure IPython on the small one. Takeaway: savings from routing are easily eaten by harness token overhead—whether you save money depends on the whole protocol, not model size.
More from coding & agent
- We tested 12 LLM classifiers — the new classifier hype is LARP — KlausCodes · 2026-10-23
- Dev builds ESP32 desk display showing Claude Code and Codex agent status via local Go helper — tonylook · 2026-10-23
- Let your agent read its own traces: AgentsView + history tool debugging workflow — dbreunig · 2026-10-23
- YC-backed Mireye launches land search API + MCP with 160M parcels for AI agents — ycombinator · 2026-10-23
- Testing 12 decision systems: the new LARP hype around classifiers — KlausCodes · 2026-10-23
- Monid launches 'Product Hunt for agents,' ranking tools by real agent calls — claudeai · 2026-10-23