How to bill clients for Claude Code and Codex usage
AI coding agents are now a real line in an agency's costs, and they are spent on specific clients' code. Here is how to decide whether to bill for them, which model to use, and how to produce numbers a client will accept.
Updated
Short answer: measure AI agent usage per client from the session logs, price it at the vendor's list rate, agree a billing model in the contract (pass-through, markup, included allowance or fixed fee), and show it as its own invoice line with the period, sessions and rate basis. Tallyhook produces those lines automatically.
Should you bill clients for AI usage?
It depends on how material the cost is. Anthropic reports that Claude Code averages about $13 per developer per active day and $150 to $250 per developer per month across enterprise deployments (source), and heavy agentic work can run far higher. For a small agency that can be a meaningful share of margin on a fixed-fee project, and it varies a lot between clients. When it varies by client, averaging it into your rate means low-usage clients subsidize high-usage ones.
Four billing models
1. Pass-through at cost
Bill the measured usage for the period with no margin. Easy to explain and hard to dispute. Best when clients are sensitive to AI costs or you want the line to feel purely transparent.
2. Pass-through with markup
Add a stated percentage to cover tooling, review and administration, as many agencies do for other third-party costs. Put the percentage in the contract and on the invoice.
3. Included allowance, then overage
Include a monthly amount of AI usage in the retainer and bill anything above it. Clients get predictability; you are protected from runaway months.
4. Fixed fee, priced with real data
Don't bill usage at all, but use measured per-client history to price the next fixed-fee project correctly. Even if you never show the number to the client, you need it.
Measuring usage per client
Whichever model you pick, the number has to be per client and reproducible. That means attributing every agent session to the repository it ran in, mapping repositories to clients, and pricing each session's tokens by model, including prompt-cache reads and writes. The tracking guide covers every way to do this. In Tallyhook it is automatic: sessions are tagged by git remote, and each client has an invoice view.
On subscription seats (Claude Max, Team, Enterprise or ChatGPT plans) there is no per-session bill. Use the list-price equivalent, the cost of the same tokens at public API rates, and label it that way. It is consistent across clients and can be checked against the published prices.
Contract wording
A clause along these lines keeps it clear (have your own counsel review it):
"We use AI coding tools in delivering the services. Usage attributable to your project is measured per session and billed monthly at the providers' published list rates [plus N%]. Each invoice includes a breakdown by period and model."
What the invoice line should show
- The period (for example, September 1 to 30)
- Number of agent sessions and, if useful, developers
- The basis: list-price equivalent at published API rates
- The list amount, the markup if any, and the billable total
- A link to a read-only breakdown, if the client wants detail
AI agent usage, September 2026 (59 sessions, list-price equivalent) $600.42 Handling (20%) $120.08 Billable $720.50
Tallyhook's invoice view produces this per client, exports it as CSV for Xero, QuickBooks or a spreadsheet, and can share a read-only report link so the client sees the breakdown without a login.
Questions
Should agencies charge clients for AI coding tools?
If AI agent usage is a material cost of delivering a client's work, it is reasonable to recover it, the same way agencies pass through hosting or stock licences. The key is to agree it in the contract up front and to bill from measured usage per client rather than an estimate.
How much markup is normal on AI usage?
There is no industry standard yet. Many agencies treat it like other pass-through costs: at cost, or with a handling margin to cover the tooling and administration. Whatever you choose, state the percentage in the contract and show it on the invoice.
How do I bill AI usage when we pay a flat subscription?
Price each client's sessions at the vendor's public list rate. That list-price equivalent measures the value of the usage consumed for that client, is reproducible from the logs, and does not depend on how your seats are bundled. Say on the invoice that it is a list-price equivalent.