Echo routes open-weight models to match Fable at one-third the inference cost

adam_rida · hn · 2026-07-24

Echo is an experiment in treating a pool of open-weight models as one system instead of picking a single model for every task.

The author evaluated models including GLM-5.2 and Kimi K2.7, then built a router that decides how much compute to spend, which models to call, and how to combine their outputs.
In early tests, Echo outperformed the best single model in the pool and roughly matched Fable’s aggregate result at about one third of the inference cost.
The remaining challenge is making better allocation and combination decisions, especially on coding and agentic tasks where quality is harder to measure.

More from Infra