A forum user posting as ggvh on the Hugging Face community forums reports that moving small, bounded decisions away from an agent’s main language model cut its calls to that model by 29.6%. The numbers are self-reported and unreviewed.
The test paired Qwen 3.8 27B with a hosted decision model called Jev across 360 SOP-Bench cases from three workflows: patient intake, referral abuse checks, and email intent. Per the post, main-model calls fell from 2,537 to 1,785, estimated cost from $6.41 to $4.70 (26.6%), and mean case latency from 8.64 to 6.94 seconds (19.7%). The cost figure is a calculation, not a bill, and falls to 24.6% once final plan preparation is counted.
ggvh discloses building Proceda, the software that runs the agent, and running the experiments. The post links public code and reports. We have not checked them; a reader should start there. The 360/360 accuracy, against 357 for Qwen alone, is not evidence of improvement, the author says.
One benchmark, one model family, one builder. The direction matches other recent reports of savings from routing small decisions away from a big model. Rerun it on your own agent logs, keeping a fallback.
Based on a post by ggvh on the Hugging Face community forums, 12 October 2026
