How do dev shops bill AI coding agent costs to clients?
Give each client project its own API key on a fixed monthly plan, and put the plan price in the quote as the AI tooling line. On Standard Compute, plans start at $19 a month and a key stops at its monthly compute budget instead of billing overage, so the most a project can cost is known before you quote. One key works in Claude Code, Codex CLI, OpenCode and Cline.
Why per-token bills break fixed-price quotes
On our gateway's coding-agent traffic, a typical active day on Claude Sonnet 5.5 costs $3.67 at list prices and a busy day $29.55, about 8 times as much. Which kind of day a project gets depends on the codebase and the task, and you find out after the work is done. On one shared OpenAI or Anthropic key, every client's usage also lands on the same bill, so splitting it per project means tagging and spreadsheets.
A key per project fixes both: each project's spend is counted on its own key, and the key's plan caps it.
Set up a key per client project
- In the dashboard sidebar, under Business, choose Create a company and name it after your shop.
- Invite the developers from People & keys with + Invite teammate.
- Choose + Key, name it after the project (for example acme-portal), leave it under the company, where new keys start, and pick a plan. Nothing is charged until an admin activates the plan.
- Open the key and, under Who can use it, choose + Add a person for each developer on the project. They can then copy it from their own dashboard. A person belongs to one team at a time but can use any number of project keys, so share keys instead of making a team per client.
- For an EU client, set the key's Compliance to Europe.
- Each developer puts the project key in that client's repository, for example in its
.envfile, so the agent uses it only there. - When the project ends, choose Cancel renewal and the key keeps working until the paid month runs out, or Revoke key to stop it at once.
Worked example: a 3-month fixed-price project
Two developers, each with a typical day on Sonnet 5.5 every working day, use about $73 each, so the project key needs about $146 of usage a month. Every plan includes a compute budget a little above its price: you pay the price, and usage past the budget is refused.
| Plan on the project key | Price per month | Compute budget | 3-month quote | Covers |
|---|---|---|---|---|
| Standard | $89 | $95 | $267 | One developer on typical days |
| Pro | $249 | $269 | $747 | Both developers, plus about 5 busy days a month |
| Pro Plus | $499 | $549 | $1,497 | Both developers, plus about 16 busy days a month |
Usage from the Agent Burn Index at Claude Sonnet 5.5 list prices, 20 working days a month; a busy day adds about $25.88. The sample is agents such as OpenCode, Kilo, Cline and Roo; Claude Code and Codex may differ, so run a week of real work on the key before you send the quote.
Which number goes in the quote
- Pro fits this project: $747 for 3 months, the same whatever the usage.
- Pay as you go costs usage plus 15%, about $168 a month at this estimate, and moves with usage. A monthly limit on the key caps it; quote the limit plus 15%.
- A key per developer per project also works: each developer gets their own budget, and one heavy week can't use up a colleague's.
EU clients: keep inference in Europe
With Compliance set to Europe, the key's requests, fallbacks included, run only at model providers' European endpoints. Claude Sonnet 5.5, Claude Opus 5.5 and GPT-5.6 are among the models available there; the rest of the catalogue is off for that key. A rule on the whole company or a team is a ceiling that keys under it can narrow.
The rule sets where inference runs. Requests still pass through our gateway and OpenRouter, which can process them outside the EU. What to tell a client is on the EU coding agent API page and our data privacy page.
Invoices and passing the cost on
- Each plan key is its own monthly subscription with its own invoice, addressed to your company with the billing address and VAT or tax ID you enter at checkout.
- Every paid invoice is emailed to the company's billing email (or its owner). The invoice shows the plan name, and Billing lists each key with its plan and renewal date, which is how you match an invoice to a project.
- You re-invoice the client: the plan price at cost as its own line, or rolled into your fixed price or day rate. If the client should pay us directly, set them up in an agency workspace instead: see the LLM API for agencies.
- Pay as you go is one prepaid balance for the whole company, so its top-up invoices cover every pay-as-you-go key. Each key's spend this month is in the People & keys card on the company Home.
Good to know
- Plans bill monthly from the day each plan starts, with no annual contract. Pay-as-you-go limits reset on the 1st of each month (UTC).
- Everyone on a shared project key draws from one budget.
- A key on a monthly plan can switch to pay as you go only after that plan ends.
The catch
The plan is a hard ceiling. If a project uses more than the plan's budget, the key stops with HTTP 402 mid-sprint until you upgrade it, and the upgrade is charged straight away, prorated for the rest of the billing period. Size the plan with room for busy days.
Where Standard Compute fits
Standard Compute fits dev shops that quote fixed-price work and want each client's AI spend fixed in advance: one company workspace, a key per client project on its own plan, Europe-only keys for EU clients and a monthly invoice per key. Creating the company is free; billing starts when the first plan is activated.
FAQ
Should a dev shop charge clients for AI coding tools?
Many fold it into the day rate or fixed price; others pass it through at cost as a separate line. Either is easier with a key per project, because each project's cost is a known plan price instead of a token bill you split afterwards.
What happens if a project key runs out of budget mid-sprint?
Requests on that key are refused and every other key keeps working. An admin can move the key to a bigger plan from its panel in People & keys, and the new budget applies at once.
Can the same developer work on several client projects?
Yes. Add them to each project's key under Who can use it, and they use the matching key in each client's repository. Their personal key stays separate for internal work.
Do we get a separate invoice per client?
You get one invoice per plan key each month, so a key per client project gives each project its own invoices, addressed to your company. For invoices addressed to the client, use an agency workspace with a workspace per client.
Which coding agents can use a project key?
Any agent that accepts an OpenAI- or Anthropic-compatible base URL, including Claude Code, Codex CLI, OpenCode, Cline, Kilo Code and Aider. The base URL is https://api.stdcmpt.com/v1, and Claude Code uses ANTHROPIC_BASE_URL=https://api.stdcmpt.com.
Does client code get stored?
We don't store, log or retain prompts or responses. They pass through the gateway in memory and go to the model provider that serves the request.