.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
Run OpenClaw on a pay-as-you-go API key for a week or two and you learn to dread the billing dashboard. The number is always higher than the work felt like it deserved.
The headline cases are wild. OpenClaw's own creator reportedly ran up around 1.3 million dollars of OpenAI tokens in a single month, something like 603 billion tokens across 7.6 million requests, and OpenAI covered it because he works there. Most of us sit much further down the scale, but the shape is identical. People share their invoices and you see a 623 dollar month, or 200 dollars gone in an afternoon, or a quick five-minute job that somehow ate 30 dollars before anyone noticed.
The cause sits in how OpenClaw works...
Every task goes to whatever frontier model you set as default, so a heartbeat check, a JSON tidy-up, a "is this spam" classification and a small PDF field grab all get billed at full reasoning rates. You pay premium prices for work a much smaller model could finish for a fraction of a cent.
This is why we partnered with Neurometric. They build and host task-specific small models that hit frontier-level quality on the jobs they're trained for, at roughly 90 percent less cost. Picking the right model for each task is handled automatically, but the models themselves are the point.
What ClawPack does
Our new partner built exactly the fix for that. Neurometric's ClawPack is a set of small models, each fine-tuned for a specific kind of task, that take the routine work off your frontier model. Classification, extraction, formatting and summaries go to the specialist built for them. Anything that needs real reasoning falls back to your own frontier model, the one you already configured. ClawPack reads each prompt and hands it to the right specialist, so you never choose manually.
It's 39 small models behind a single OpenAI-compatible model ID, covering six areas: legal, finance, coding, sales, support and marketing. The part that matters if you already have a working setup is that ClawPack doesn't replace your default model. It sits next to it, so nothing you built changes.
Neurometric says this drops frontier API calls by up to 90 percent on a typical agent. The savings hold up, the exact figure just depends on how chatty your agent is and how much of its day is genuine reasoning rather than busywork.
Running ClawPack on your LumaDock VPS
Most of you reading this already run OpenClaw on a LumaDock box, and that's the good news. ClawPack runs on the same OpenClaw VPS you already have, so there's no new server and no migration. There's a free tier of 100 million tokens a month with no credit card, which is enough to test it honestly before you decide anything.
You can follow the full ClawPack setup tutorial and have it routing today.
