The real cost of a working agent team is roughly 10 shekels a day, and most of that is a flat monthly subscription, plus a few shekels on automations you pay for by usage. That number is almost always smaller and more boring than people imagine, because what really decides whether it pays off is your time, far more than the tokens.
As the person who runs operations here, this is what the monthly bill looks like once you break it into parts:
- The flat subscription. Most of the cost comes from one monthly plan for the model, a fixed amount that doesn't change whether you ran one session on it or a hundred. That's the base, and it's almost the whole story.
- Metered automations. You only pay by volume for things that run on their own in the background (image generation, speech to text, search), and that comes out to a few small shekels a day. One time we forgot an automation left running on empty, and paid for it for three days before we caught it, so now we have a short weekly check.
- Everything else. The files, the memory, the messages between agents: zero shekels. It's just text on your disk, and it costs nothing.
The number stays low thanks to one simple thing: discipline. Every agent reads only what its task actually needs, and doesn't load the whole brain for every small question. When you hold that line, the same model is plenty for the entire business at roughly 10 shekels a day (most people find they pay more for one streaming subscription they forgot to cancel).
So instead of asking how much it costs, the right question is how many hours it hands back to you. A team that costs less than a sandwich a day, and saves you two hours of routine work, has already done the math for you.
A prompt, on the house
Run yourself a monthly cost check in three lines:
1. Flat subscription -> how much do I pay for the model plan per month?
2. By usage -> which automations run in the background and bill by volume? (mark each one)
3. Time -> how many hours of routine work does the team save me per week?
If line 3 is worth more than lines 1+2 combined, it pays off.
If not, shut down one automation you haven't touched in a month, and measure again.Start small: one subscription, one automation that genuinely saves you time. You add the rest only when something hurts enough to justify it.
