An Alibaba Cloud post reports a routing experiment built on HiRoute, the engine the same team open-sourced this week. The assignment was a messaging-system design review: research six candidate technologies, check the official documentation, and deliver a memo on one decision — how a system avoids processing the same event twice, or dropping it, when it fails midway and retries.
Three configurations ran the same task with the same source material: all GPT-6 Astra, all Qwen-3.8 Flash, and a split in which Qwen produced the six research deliverables and Astra wrote the decision memo. Two of three runs cleared the original whole-delivery gate. Across those two, research-check accuracy was 99.03% for the split and 99.72% for all-Astra; critical decisions without major errors were 3/3 for the split, 3/3 for all-Astra and 1/3 for all-Qwen. Whole-task cost for the split was 91.39%–92.01% below all-Astra, counted at API-equivalent prices and including failed attempts and routing evaluation.
A second case, adding synchronous and asynchronous JSON stream iteration to HTTPX over roughly 32 minutes, exercised the failure path: when the economy model kept researching without moving into implementation, the router upgraded the model at a context handoff, and the result passed 343 independent checks with no manual model or prompt change.
The post defines same quality as meeting the same delivery standard, not as matching scores.
Our takeThe load-bearing number is 1/3, not 92%. The gap opened on the judgment step, where the weaker model treated a guarantee inside one component as a guarantee for the whole design. Two qualifying runs is thin evidence, and the delivery gate was written by the same team shipping the router — so read 92% as an upper bound on a favourable task shape, not a general rate.