KeelenAnswer · what it costs
What does it cost to run an AI coding agent on a real project?
Updated 11 September 2026
The bill has two halves. One is what the tool charges you, and one is what the model uses. The second half surprises people. Input tokens drive it, and the amount of code written does not, because an agent reads the same files again on every step of a task. We measured that. Across 12 controlled runs of the same specification, Claude Code, OpenAI Codex, Moonshot Kimi each recorded a median of 3 merged pull requests. Their use of input tokens differed by about 4.7 times between the heaviest and the lightest. Those token figures are a consumption snapshot. They are not billed cost, accepted work, or equal quality. Keelen charges a flat monthly fee (Indie $29, Operator $79, Agency $299). Your coding-engine token usage bills to your own provider account at cost, with no markup added by us.
Those token figures measure our loop driving each engine. The projects were small and newly created. The figures are medians, and the sample size sits beside each one. They are not a general claim about what any vendor's model costs in someone else's tool.
What to do, in order
Separate the platform fee from the model usage
Some products bundle model usage into one price and some run on credentials you connect. Bundled looks simpler and hides the variable. Connected looks like two bills and lets you see, and change, the part that actually moves.
Meter input tokens
Output is the code, and it is small. Input is the context: the files read, read again, and sent again on every step. A modest feature can use millions of input tokens and a few thousand output tokens. An estimate built on lines of code will be wrong by an order of magnitude.
Price the unit you actually care about
Track provider bills, platform fees, and the review effort on your own repository. A merged pull request count is a weak signal. It does not show billed cost, accepted work, or quality.
Check what your existing subscription already covers
Suppose you already pay for a coding plan. Running an agent on those credentials uses an allowance you have bought. It does not add a metered bill. A heavier engine then costs you headroom instead of money. That is a different constraint.
Cap the autonomy you cannot supervise
The costly failure is a loop that retries a task it will never finish. That loop can burn more than any single hard task. Look for retry budgets on each task and a daily ceiling. Look for a project that pauses and asks for help. That beats burning quota against a dead end.
Why input tokens dominate
An agent working on a real repository spends most of its budget reading. It opens files to find the relevant code, then reads them again as the task develops. It sends that context again on every step of the conversation. The code it writes is a small part of the traffic. So two tools can make the same change for very different money. The size of the diff predicts almost nothing about the cost.
What we measured, and the honest scope of it
12 controlled runs between 2026-07-20 and 2026-08-10. Every engine got the same freshly written specification. They also got the same starter repository and the same merge gates. The projects were Python libraries and web apps. They were small and newly created. Every figure is a median. Its sample size sits beside it. The sample is small, so treat all of it as directional. We do not rank the engines on quality. Twelve runs cannot support that claim about anybody's product.
Whose account pays for the tokens
Keelen runs on credentials you connect. Anthropic Claude works with a Claude Code subscription login or an API key. No Max plan is required. You can also connect OpenAI Codex, GLM, Kimi, or xAI Grok. Coding-engine token use bills to your own provider account at cost. Keelen charges one flat monthly fee and adds no markup.
What stops a run from burning money on a dead end
Failed runs fall into 30 kinds. Small infrastructure blips and provider limits retry on a budget. They never count against your work. A real dead end becomes one plain English card in your Needs-you queue, with a recommended action. The project pauses instead of burning quota.
When Keelen is not the answer
- You want one bill with model usage included. Keelen does not resell tokens. You will see a provider bill or a provider allowance. That comes on top of the subscription.
- You have a handful of small tasks a month. A chat subscription you already own is cheaper than any continuous system. That is the honest answer.
- You need a firm quote before you start. Usage depends on your codebase and your work. Any number quoted in advance, here or anywhere else, is an estimate.
Hand the work to a loop
Connect a repository and write what you want in plain language. Review the tested pull requests that come back.
FAQ
What does it cost to run an AI coding agent every month?
Budget the platform fee plus your own model usage. The platform fee on Keelen starts at $29 a month. Model usage depends on your codebase and how much work you queue. Coding-engine use bills to your own provider account at cost. The engine you pick moves that second number more than anything else you control.
Why is the token bill so much larger than the code produced?
Reading costs more than writing. An agent reads and sends the same context again on every step of a task. So a change of a few hundred lines can use millions of input tokens. Output tokens are the actual code, and they are a small part of the total.
Is a cheaper engine worse?
The benchmark does not answer that. It records dated consumption. It also records merged pull request medians. That number is 3 in this snapshot. It does not measure equal quality, customer acceptance, or billed cost. Test a provider on your own repository. Check its review process too.
Can I use a coding subscription I already pay for?
Keelen runs on credentials you connect rather than on tokens we resell. Connect Claude Code, Codex, GLM, Kimi, or Grok. Grok accepts a SuperGrok sign in or xAI API key. What your plan permits is set by that provider's own terms, so check them before you rely on it.