Solibra's Sovereign router learns model choice without labels

Solibra built Sovereign, a router that sits between an application and model providers and learns, per request, which model to use by grading its own finished work — with no answer key and no prior knowledge of the models. It speaks the same APIs as the providers, so adopting it is a configuration change rather than a rewrite, though each request adds a routing decision of about half a second. Across eight public benchmarks against Claude Opus 5.5 and Claude Haiku 5.5, Solibra reports Sovereign spent 43% to 81% less than Opus on every benchmark, matched or beat Opus accuracy on USACO, LiveCodeBench and SATBench, and was more accurate than Haiku everywhere, by up to 55 points on USACO. On USACO its accuracy rose from 88% over the first 25 tasks to 97.7% by the end, passing Opus around task 150, and its grader agreed with the hidden answer key 95% of the time. These figures come from single runs, with cost measured as model spend at list prices, Solibra says.

สร้างจากรายงาน · อัปเดต 11 ชม. ที่แล้ว

ยังมีการเฝ้าดูต่อเนื่องไม่พอจะวาดแนวโน้ม