Leaderboard
ServeLearnBench leaderboard
Avg. Hidden averages the nine Hidden-Dependent settings and Avg. Fully Spec. the six Retail and Banking Fully Specified settings, each setting counted once. API cost is the total over nine settings, serving stream, learning work and test included.
MethodPrimeCHRAGMem0SkillOpt
ModelGLM-5.3 FlashDS-V4.1 FlashGLM-5.3Kimi K3Opus 5GPT-5.6 Terra
Results snapshot 2026-09-25. One run per cell. Prime was not run with Opus 5 and GPT-5.6 Terra because of its cost.
The frontier view plots Avg. Hidden against the cost of one held-out test task; the staircase is the Pareto frontier.
Breakdown
Hidden score by setting
Each cell is a setting score on the held-out test tasks. Darker is better; hover for the value.