跪拜 Guibai
← All articles
Backend · Programming · Artificial Intelligence

OpenSquilla Routes Every AI Task to the Cheapest Model That Can Finish It

By cxuanAI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

No single developer can keep up with near-monthly releases from multiple model vendors, and the gap between an oversized model and an adequate one shows up directly in every token bill. OpenSquilla moves that benchmark work to a monthly lab test that ships the cheapest sufficient config straight to the client, so per-task spend stays near the floor without anyone tracking model churn.

Summary

Model releases from DeepSeek, GLM, Qwen and Kimi land nearly every month, each vendor claiming another leap in code ability. Manually testing which model wins which task, and which one costs least for that job, stops being feasible early on. OpenSquilla, TokenRhythm's open-source agent framework, answers with a local router that drops each task into one of four difficulty tiers (C0-C3) and dispatches it to a single model. One request calls one model, never a parallel candidate set, so token spend stays near the minimum. A task keeps its tier across continuations; only a redo unlocks a stronger model.

TokenRhythm's lab re-benchmarks model combinations every month on Claw-SWE-Bench and AppWorld, weighing quality against spend, and selects the config on the Pareto frontier. November's first list pairs Qwen3.8-Flash with C0, DeepSeek-V4.1-Flash with C1, GLM-5.3 with C2, and Kimi K3 with C3. Against older configs, the November set improves code-repair score by 32.41% while cutting per-task generation cost 22.10%; AppWorld gains 9.39% at a 2.96% cost increase, a stated trade-off for long multi-step cross-app tasks. Configs update independently of the client at startup and every 24 hours, and apply only after the user confirms the diff.

A survey of mainstream AI coding tools rendered as a Chinese interactive dashboard ran through smart routing and landed on Kimi K3. Mac and Windows desktop clients cover casual users, uv installs without a local Python, Docker images ship for amd64 and arm64, and source builds track the main branch. Model access runs through a TokenRhythm account, with 18 yuan of trial credit after real-name verification.

Takeaways
— A local router assigns every task to one of four tiers (C0-C3) by difficulty and selects the model for that tier.
— Each request invokes exactly one model; no parallel multi-model runs, which avoids multiplying token costs.
— Repeated steps of the same task keep the original tier; a redo is the only trigger to move up a tier.
— November's first monthly config: Qwen3.8-Flash for C0, DeepSeek-V4.1-Flash for C1, GLM-5.3 for C2, Kimi K3 for C3.
— On Claw-SWE-Bench, the November config scored +32.41% with per-task generation cost down 22.10% vs prior configs.
— On AppWorld, the November config scored +9.39% with per-task cost up 2.96%, a stated trade-off for long cross-app tasks.
— Recommended configs update separately from the client, checked at startup and every 24 hours, applied only after user confirmation.
— Install routes: Mac/Windows desktop client, uv-based terminal install without Python, source build requiring Git, Git LFS, Node.js 22.12+ and uv, plus a Docker image for amd64 and arm64.
— Model access requires a TokenRhythm account; real-name verification grants 18 yuan of trial credit.
Conclusions

Enforcing one model per request is the cost lever that makes routing worth having; schemes that run several candidates per task pick a better answer but multiply token spend, and that extra spend defeats the purpose of a cost-optimizing router.

The AppWorld result (+9.39% score at +2.96% cost) is a deliberate trade-off: cross-app tasks with long step chains need a pricier model to lift success rate, so the cheapest-model rule accepts occasional cost increases where the task demands it.

Requiring user confirmation before applying a fresh monthly config turns the benchmark refresh into a visible operation, a trust mechanism that prevents silent model switches from changing either cost or behavior under the user.

Concepts & terms
Smart routing (智能路由)
Automatic per-task model selection that replaces a fixed model; OpenSquilla's router picks the model from a difficulty tier, aiming for the cheapest model that can complete the task.
C0-C3 difficulty tiers
Four levels a local router assigns to each task, from simplest (C0) to most complex (C3); each tier maps to a specific model from the monthly recommended config.
Pareto frontier
In the monthly benchmark, the selected config is the one no other tested option beats on both cost and quality; it offers the best result at the same cost and the lowest cost at the same result.
Claw-SWE-Bench
A benchmark for code-repair tasks used in the monthly config tests.
AppWorld
A benchmark for cross-application transaction tasks used in the monthly config tests.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗