Coding agent costs: model, compute, and review

Build a ledger for each delegated task

Give each task an identifier that connects its sessions, model charges, and final pull request. Record the result even when no change ships. Failed attempts used capacity and someone had to inspect them. Omitting them makes a workflow appear cheaper than the queue your team actually runs.

Cost	What to capture	Common accounting mistake
Platform	Subscription or usage fee and allocation method	Adding included usage again as a separate expense
Models	Actual provider charge, model, and retries	Counting your own API invoice as free usage
Execution	Any separately billed VM, storage, browser, or network use	Assuming every provider bills compute separately
Engineering time	Setup, supervision, review, repair, and environment upkeep	Counting only the final successful review

Keep elapsed time in a separate column. A task that runs unattended for an hour has a different staff cost from one that needs an engineer watching for an hour. Both may occupy the same execution capacity.

Read the billing unit before comparing prices

A credit balance, a model token charge, and a seat license measure different things. Identify what a plan includes, what creates an overage, and which party sends the invoice. Check the renewal policy before treating unused allowances as savings.

Uno's pricing page lists monthly plans with an organization-wide credit balance. It states that monthly credits expire at renewal, actual Uno AI Gateway costs are deducted from credits, and your own model keys and subscriptions are not. With your own provider, account for that provider's bill or subscription separately.

Avoid adding a subscription payment and the same included model usage as two cash expenses. For internal attribution, you can allocate the subscription across tasks and report consumed credits beside it. Mark credits as an allocation measure so finance can reconcile the cash total with the actual invoices.

Cursor says Cloud Agents consume plan usage at the selected model's rate. Legacy request-based plans have different billing, and larger context windows can increase cost. Check the plan's billing unit before comparing it with another service. Cursor Cloud Agent billing.

A hypothetical cost-per-change calculation

Suppose a team delegates ten similar fixes. Seven become accepted changes. Across all ten tasks, the team allocates $20 of platform spending, pays $30 for model usage outside that allocation, and incurs $10 of separately billed execution costs. Total cash cost is $60. These numbers are illustrative, not Uno prices or measured results.

The team also spends two hours reviewing successful changes, one hour investigating rejected work, and one hour setting up environments. At a hypothetical internal rate of $75 per hour, that is $300 of staff time. The cost per accepted change is:

Cash cost:             $60 / 7 = $8.57 per accepted change
Staff time:          4 hours / 7 = 34.3 minutes per accepted change
Attributed full cost: $360 / 7 = $51.43 per accepted change

Keep all three numbers. Cash cost supports a spending budget; minutes show the demand on reviewers; attributed cost lets you compare with another workflow using the same internal rate. A different staff rate changes the last number without changing the actual provider bills.

In a longer trial, spread one-time setup over the period that benefits from it. Report the initial investment and recurring upkeep separately. Otherwise the first week looks unusually expensive and later weeks conceal the work needed to keep environments usable.

Put a budget around retries and parallel work

Decide what ends an attempt before starting it. A missing credential needs environment repair. A reproducible test failure needs diagnosis. An ambiguous task needs a decision from its owner. Repeating the same instructions without resolving the blocker can add charges without producing a better change.

Set an initial task budget and a point where a person reviews further spending. Include a rule for abandoned sessions and the infrastructure they leave behind. When a provider reports usage late, label the task's cost pending until the charge settles.

Parallel execution needs reviewer capacity too. If four agents finish while one reviewer is already busy, their changes wait. Branches may need updates and previously green checks may need another run. Increase concurrency after measuring queue age and correction work. The parallel coding agents guide covers how to divide tasks before adding more simultaneous sessions.

Use a small pilot to find your useful task types

Group tasks by a practical category such as dependency updates, UI bugs, or test coverage. Keep a consistent acceptance policy within each group. A documentation fix and an authentication change demand different review, so one blended average gives little guidance about what to delegate next.

Record accepted, rejected, blocked, and still-pending tasks with their costs.
Track human minutes and the time from delegation to acceptance.
Revisit accepted changes after release and record repairs or reversions.
Compare against similar work in your current process using the same accounting rules.

Resist selecting a winner from one unusually easy task. Keep the individual results alongside the average so a costly failure stays visible. Write down why the team accepted each change; a higher acceptance rate is useful only if the acceptance checks still protect the behavior you care about.

Before expanding the budget, identify the task categories with acceptable cost, review effort, and release results. Use the verification guide to keep those criteria explicit as the queue grows.