Model routing nabs 60% savings on financial benchmarks

This benchmark proves that "model routing" isn't just for general chat—it's a critical cost-optimizer for domain-specific RAG pipelines.

• –Routing sub-questions within complex documents (like 10-K filings) is the "killer app" for Haiku-style small models, yielding 58% savings even on reasoning-heavy datasets.
• –The failure to route long-form transcripts (ECTSum) suggests context window density and prompt length are still the primary bottlenecks for router scoring.
• –Using open-weights models like Qwen 3.5 and Gemma 3 for mid-tier tasks yielded up to 89% savings, highlighting the maturity of 27B-class models in specialized verticals.
• –The "Flexible" strategy outperforming "Intra-provider" routing in sentiment tasks suggests that proprietary models have a diminishing lead in basic classification compared to self-hosted alternatives.