Less than one week — that's our measured answer, not an estimate. Anthropic offered Spotlight Links a $100 Claude Code credit; servicing five active clients, it lasted six days and bought roughly ten hours of dedicated agentic work, an effective rate of about $10 per working hour. We bought $50 more before the week was out.
We publish this because when we searched for real agency burn data before spending, none existed. The full breakdown, with charts, is on the blog.
Why it goes faster than the token prices suggest
Per-token pricing looks tiny — a few dollars per million tokens reads like the meter will never move. But agentic coding isn't chat. In a Claude Code session the model reads files, edits them, runs builds, ingests the output, and re-reasons with full context on every step. One focused hour moves tens of millions of tokens. The meaningful unit is dollars per hour of work done, and ours is ~$10.
What should an agency actually budget?
From our logs, scaled linearly and rounded honestly:
| Load | Monthly credit budget |
|---|---|
| 1–2 clients, part-time | $50–$100 |
| ~5 clients, active servicing | $200–$400 |
| ~10 clients, active servicing | ~$500 |
A per-client rule of thumb: $40–$80 of model spend per actively serviced client per month, higher in build-heavy months. Treat it as cost of goods sold — the spend maps directly to billable output.
Is that expensive?
The hours it buys produce work that has always billed at hundreds of dollars per hour — full-stack builds, deployments, search architecture. Ten dollars for an hour of that output is the widest margin in the stack. Our reorder decision took about thirty seconds, which is the most honest product review we can give.
What burns credits fastest
- Agentic sessions on real repos — the big one, by an order of magnitude over chat
- Deployment work — logs, DNS, build output, all interleaved into context
- New-client audits — a month of reading compressed into days
- Vague prompting — wrong turns cost the same as right ones; scoping discipline is the cheapest optimization there is
Should you wait for it to get cheaper?
It will get cheaper — AI price-performance improves at least 30% per year, and that curve has a decade or more of runway. The $500/month budget of today buys meaningfully more output every year, making credits the only business expense that deflates while improving.
But the price falls for everyone at once, including your competitors. The compounding advantage goes to whoever is already fluent when it does. Waiting saves you dollars and costs you the head start.
Caveats
This is one agency's measured week, not a benchmark. Your mix of build versus advisory work changes everything, and model pricing changes over time — verify current rates before budgeting from our numbers.
Curious what those hours produce on the client side? Book a free consultation — the first four booked from this site are free.