What AI agents really cost to run — and how we keep them affordable

Most of what you read about AI agents is either hype or fear. This post is neither — it's our meter readings. Nova runs autonomous agents that do real work for local businesses: updating websites, writing blog posts, reviewing search performance, posting to social media, replying on Google. Every one of those runs costs real money, and because our customers pay a flat monthly price, that cost is our problem to engineer away — not yours.

We build in public, so here's what six weeks of running agents actually looked like, including the two days that went wrong, and the five lessons that came out of it.

What one agent run costs

When an agent updates a client's website, a frontier-class AI model reads the site's code, plans the change, makes the edits, checks its own work, and publishes. Across our ledger, a single run costs anywhere from about 8¢ (₹7) to $1.40 (₹123) depending on how much reading and checking the task needs. A typical day across all our businesses lands between $0 and $5.

Two days did not.

Those two red bars are the honest part of this post.

Lesson 1 — The expensive model only sits where work gets produced

The single biggest lever is architectural. Every message your receptionist answers, every chat you have with Nova, every summary you receive — those run on fast, lightweight models that cost a fraction of a rupee per message. The frontier-class model is reserved for the seats where work products come out the other end: editing your site, writing long-form content, analysing your search data.

Cheap models in the busy seats, the expensive model in the productive seat. Get this backwards and no amount of optimisation saves you.

Lesson 2 — Retries, not big jobs, are the silent budget-killer

That second red bar is the night of 17 September. A change had been approved and one step of it couldn't be finished automatically. The system did what it was built to do — try again. And again. Between 12:52 am and 2:08 am it launched 13 runs in 76 minutes, 8 of which failed, spending $11.75 (about ₹1,030) while everyone slept.

No single run was expensive. The loop was expensive. Our fixes, live the same day:

• A hard retry ceiling. Two attempts per piece of work. After that, it's routed to a human — never a third robot attempt.
• One run at a time per business. Work is serialized, so a stuck task queues instead of multiplying.

Lesson 3 — Hard caps beat good intentions

Every guardrail above is logic, and logic can have bugs — ours did, that's how the night happened. So underneath the logic there is now a number that doesn't negotiate: a daily spend cap per business — $5 (about ₹440) by default. When the day's pipeline work reaches the cap, agents stop and resume tomorrow.

Two details we're proud of:

• The meter fails closed. If the spend-tracking itself ever errors, agents treat the cap as reached. When you can't read the meter, you stop the car.
• Hitting the cap never loses work. Half-finished changes are saved on a side branch and picked up next run. You paid for those tokens once; we don't burn them twice.

Lesson 4 — Remembering is ~10× cheaper than re-reading

Inside a single run, an agent works in rounds: read, think, act, look at the result, think again. The naive way, every round re-reads the entire conversation so far at full price — so cost grows with the square of the rounds. Model providers offer a fix: previously-seen context can be cached, and cached re-reads are billed at roughly a tenth of the price.

We restructured our agent loop so the frozen part of the conversation stays byte-for-byte identical between rounds (any change breaks the cache), and we project this cuts our model bill for agent runs by roughly half. We're measuring it in production now rather than taking our own projection's word for it.

The same idea applies outside the model too: our agents keep a read-cache ("you already read this file, nothing changed") and a verify budget (two checks per page, not nine). Agents are diligent to a fault — left alone, one of ours re-verified the same URL until we gave it a budget.

Lesson 5 — Everything gets a budget, because agents don't get tired

A human writer gets tired and stops. An agent will happily work — and bill — forever. So every dimension has a ceiling: 40 working rounds per run, a token backstop per run, two verifications per page, two retries per task, $5 per business per day. An agent without limits isn't autonomous; it's a meter without a dial.

What this means for you

• Your flat price stays flat because the engineering above, not fine print, keeps unit costs down.
• A runaway loop can cost us one bad night; it can never run up your bill.
• Cost controls never silently eat your work — capped runs park their progress and continue tomorrow, and anything the robots can't finish in two attempts reaches a human.

That's the deal we think AI agents should come with. If you run a clinic, a store, or a practice and want an AI team with its own seatbelts, start with Nova on Telegram or see NovaHub.