Save 80% on Token Costs: The Accounting of a Routing System

Conclusion first: 54,184 vs 10,875

Processing the same 50 real business requests in two ways:

• All direct connections to the cloud LLM: Consumed 54,184 Tokens
style="padding-left:20px;margin:12px 0;">
• Using the MTClaw three-layer routing: Consumed 10,875 Tokens

An 80% saving. Not a theoretical estimate, but a same-baseline actual test (2026-08-02, statistics from the /api/mtclaw-metrics endpoint).

Where the Savings Come From

Routing distribution for the 50 requests:

Three facts:

1. 76% of requests consumed zero cloud Tokens (L1 zero Token + L2 local inference);

2. High-frequency scenarios (inspection, alerting, daily reports, rule-based decisions) naturally fall into L1/L2—these are precisely the scenarios with the highest call volume;

3. Cloud API pricing is uncontrollable. Our own development logs contain records of a DeepSeek account being suspended due to a 402 CNY overdue balance, forcing a switch to another cloud. The edge-side solution also zeroes out this pricing risk.

How to Calculate Your Own Savings

We won't give you a ready-made figure like "Save X ten-thousands over 3 years"—different enterprises have vastly different call volumes. Plug your numbers into this formula:

Monthly savings = Your monthly Token consumption × 80% × Your unit price. Then subtract the one-time investment for edge-side servers (tens of thousands CNY, depending on NPU configuration). The larger the call volume and the more sensitive the data, the steeper this savings curve becomes.

Boundaries

The 80% figure is from a test sample of 50 requests; the L3 proportion in your business might be higher (e.g., long-text generation scenarios). However, the direction remains unchanged: for any high-frequency, deterministic tasks, the rule layer and edge layer can intercept the flow.

BossAgents: 71 digital employees, privately deployed on the edge, transforming the cost curve from linear growth to nearly flat.

All data in this article can be traced in the BossAgents technical documentation repository · BossAgents · Private Intelligent Agents for the Boss