What AI actually costs.
Production cost usually lands 5–20× the pricing-page number. Answer six questions about your workload and see your own estimate (inference, infra, evals, and engineering) broken out honestly.
Full breakdown, assumptions & a downloadable summary: enter your email to unlock.
Helps us send you relevant follow-ups. Skip any time.
| Inference | — |
|---|---|
| Infra | — |
| Evals | — |
| Engineering & maintenance | — |
| Monthly total | — |
| Annual | — |
Typical range across production teams: 5–20× sticker.
What this measures#
The calculator builds a monthly number from the four lines listed below. Inference starts from a base token count for the workload, then grows with context and retries before it hits a per-tier rate. The other three lines add fixed and workload-scaled overhead on top, so the total tracks actual usage instead of the list price alone. At low volume, inference is usually the smallest of the four. Infra and engineering tend to dominate a small pilot, and inference only takes over once real traffic arrives.
The four cost lines
Each line answers a different question about the workload. None of them is optional; a number that drops one of the four is not a production estimate.
| Line | What feeds it |
|---|---|
| Inference | Call volume priced at a blended rate for the model tier picked. Context growth and retry overhead are added first, on top of the clean token count. Most calculators stop here; this tool treats it as a floor. |
| Infra | A fixed monthly floor for the stage the workload is in, plus a workload-specific add-on and a small per-call logging charge. A prototype carries almost none of this. A production system carries all of it. |
| Evals | A fixed monthly eval-suite cost for the stage, plus a slice of the inference spend that gets re-run through evals on every check. Skipping this line does not remove the cost. It moves the cost into production incidents instead. |
| Engineering | A loaded engineer cost, scaled by the stage maintenance load and the workload complexity. Team size scales it further. This is usually the line a first estimate forgets entirely. |
By workload
Every workload starts from a different clean token count, then carries its own real-world overhead. The retry line is the same across every workload; the studio applies it uniformly rather than guessing which workload retries more. A chat workload grows the most from re-sent history. An agent workload grows the most from long tool loops. The infra column compares each workload against the classification workload, which carries the cheapest add-on of any workload here. It does not compare against the stage floor.
| Workload | Base tokens per call | Context growth | Retry overhead | Infra add-on vs. lowest | Maintenance complexity |
|---|---|---|---|---|---|
| Chat / assistant | 1,200 | 1.8× | 1.3× | 1.33× | 0.9× |
| RAG / search | 3,500 | 1.35× | 1.3× | 6.67× | 1.1× |
| Document pipeline | 7,000 | 1.2× | 1.3× | 5× | 1.2× |
| Agents | 15,000 | 2.4× | 1.3× | 10× | 1.4× |
| Classification | 700 | 1.05× | 1.3× | 1× | 0.8× |
Team size and the multiple
These two adjustments sit outside the per-workload table. They describe the team running the workload, not the workload itself. Both apply on top of every row above.
| Adjustment | What it does |
|---|---|
| Team factor | Adds 1% to the engineering line for every person on the team. The add-on stops growing at 20%, once the team reaches 20 people. |
| Sticker vs. real | The calculator typically lands between 5× and 20× the sticker inference price once infra, evals and engineering are counted. |
What this does not do This is a planning estimate, not a quote. It does not know negotiated rates or how efficient a given team already is at running AI in production. It also has no view into infra a team already owns. Use the multiple to sanity-check a budget, then get the real number from the people who will build it. A number this shape is meant to open a conversation with finance, not close one. Bring the assumptions along with the total, since the total on its own invites the wrong argument. Two teams can run the same workload at the same volume. Their own infra and their own team already differ, so the real number differs too. Move the sliders and watch which line moves the most. That line is where a real conversation about the build should start.
Want the real number, not just an estimate? #
This calculator gives you a directional estimate from typical production overhead. A Ship Audit gives you the actual number for your workload, plus what's driving it.