CODING AGENT API COST GUIDE
Estimate the work before you optimize the bill.
A coding agent request is more than a prompt. Separate input, output, cached context, retries, and tool loops so your estimate can explain where the budget goes.
Built around a stable OpenAI-compatible request shape.
A
coding agent can spend more on context than on the answer. It can retry a request, call a tool, or carry a large repository slice into every turn. A useful estimate starts with the work pattern, then applies the rate inputs.
THE BASIC MODEL
estimated cost = (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate)
Add cache reads, cache writes, retries, and any provider-specific terms as separate lines when they apply.
STEP 01
Count the work you send
Measure representative requests rather than a single tiny prompt. Include system instructions, repository context, tool results, and the user task.
STEP 02
Count the work you receive
Output includes explanations, patches, structured tool calls, and the extra turns created when an agent checks its own work.
STEP 03
Separate retries
A timeout or rate limit can create a second request. Keep that traffic visible so reliability work does not look like a mysterious rate increase.
STEP 04
Compare completed tasks
Cost per request is a weak metric on its own. Track cost per successful change, reviewable patch, or completed migration.
| Question | Evidence |
|---|---|
| How much context repeats? | Token counts by turn and cache behavior |
| How many tool loops finish? | Request trace or client log |
| What counts as success? | Tests, review outcome, or task completion |
| What changes after a route switch? | Same task set, new model/config, comparable window |
TRY YOUR NUMBERS
Use the planning calculator.
Replace the illustrative inputs with rates from the account catalog and a trace from your own workload.
WORKED EXAMPLE
Show the assumptions beside the total
This example is deliberately illustrative. Replace every number with values from your account catalog and a trace from your workload.
| Variable | Example | Calculation |
|---|---|---|
| Input tokens | 18,000 | 18,000 ÷ 1,000,000 × input rate |
| Output tokens | 2,200 | 2,200 ÷ 1,000,000 × output rate |
| Cache terms | Not included | Add only when the provider publishes the unit |
| Retries | Not included | Multiply by observed attempts per task |
Reduce repeated context
Trim instructions that do not change the task, scope repository context to the files the agent needs, and measure whether caching actually reduces the billed input path.
Reduce unproductive loops
Give tools clear schemas, add a stop condition, and make test output concise enough for the next turn to use.
Compare successful work
Track cost per accepted patch or completed migration. A cheap request that needs manual repair can cost more than a larger request that finishes cleanly.
Keep a change log
When a rate, model, prompt, or retry policy changes, write down the date and the reason. Otherwise a monthly total is hard to explain.
OPEN QUESTION
The right unit may be a completed engineering task
Token pricing is necessary for an estimate, but teams usually buy outcomes: a reviewed patch, a passing migration, or a resolved incident. Keep token math visible while you learn which task-level measure predicts value for your workflow.
That measure can change as the agent gains tools or the repository grows. Revisit it when the request shape changes.