What running AI agents actually costs (the part nobody publishes)
Plan prices hide the real question: what does an hour of agent work cost, all in? We bill infrastructure at cost on an itemized bill, so here are the actual numbers: what a running agent costs per hour, what it costs paused, and where the money really goes.
Every AI coding tool advertises a plan price. The number that actually decides whether you can run agents all day is different: what does an hour of agent work cost, all in, including the machine it runs on? Almost nobody publishes that. We can, because our billing itemizes every resource separately.
The two line items
Running an AI agent has exactly two costs: the machine it runs on and the tokens it consumes.
**The machine.** An agent needs a persistent cloud machine: CPU, RAM, storage, network. Ours run up to 16 cores and 64 GB RAM, billed per hour at AWS-equivalent rates with no markup. A mid-size agent machine (4 cores, 16 GB) costs about the same per hour as a mid-tier cloud VM, because that is what it is.
**The tokens.** Every time the agent thinks, writes code, or reviews a diff, it burns LLM tokens. We pass these through at the AI provider's cost with zero markup, so an agent on Claude costs what Claude costs.
What this means in practice
The surprising part is how small the machine cost is next to the token cost for real work. A machine paused between tasks costs pennies (storage only). A machine running all day costs a few dollars. The tokens for a feature that takes an agent two hours of thinking and coding can cost more than the machine did for the whole day.
This is the opposite of what plan pricing implies. Plans price the seat, and seats are cheap. Work is where the money is, which is why "unlimited for $20/month" either means heavy rationing or heavy losses.
Why itemized matters
When your bill shows machine hours and token spend separately, you can answer questions plan pricing hides:
- Was that expensive task expensive because it ran long, or because it burned tokens thinking?
- Is it cheaper to run a smaller machine for longer, or a bigger machine for shorter?
- What did last week's feature actually cost to build?
We think every AI agent platform will end up here: itemized bills at cost, because the moment agents do real work, "what did it cost" becomes a question someone asks. [Our pricing](/pricing) works this way today, with a [cost estimator](/pricing) you can play with before signing up.
The comparison trap
If you are comparing AI coding tools, watch for the pricing model, not the price. Plans with quotas are prepaying for work you may not do. Pay-while-running means you only pay when the agent works. For a team running agents every day, the itemized model usually wins; for occasional use, a small plan is fine.
The fuller comparisons: [JackHamr vs Cursor](/compare/cursor), [JackHamr vs Claude Code](/compare/claude-code), [JackHamr vs Devin](/compare/devin).
More releases
Your first JackHamr agent: from signup to a working agent in about two minutes
A step-by-step first run: create the agent, give it a job, approve its plan, and watch it work on its own cloud machine. No installs, nothing to configure upfront.
2 min read
Devin Cloud vs Devin Desktop: which do you actually need?
Cognition now ships two Devis: an autonomous cloud agent that takes tickets, and an AI editor you code in. They solve different problems, and picking wrong costs either control or hours. Here is how to tell which one your work needs.
2 min read
Focus: be the manager your agents need, without watching their screens
Focus turns managing AI agents into something like managing people: decisions arrive as cards, you answer in a tap, and the rest of the app dims so one thing gets your attention at a time.
3 min read