AI project total cost
The calculator vendors don't give you. Token prices are the headline; the bill is the stack. Fill what applies to your project, leave the rest at zero.
The formula
Token prices: daily automatic pull from OpenRouter's public API. Voice default bundles hosted STT (~$0.005/min), neural TTS (~$0.015/min) and US inbound telephony (~$0.015/min) — edit to your vendors. Data as of . Prompt-caching and batch discounts deliberately excluded (conservative case).
Token pricing is the headline; the stack is the bill
Every provider publishes a per-token price, and almost every first AI budget is built from it. That budget is wrong, not because the token price is misquoted but because tokens are not the whole product. A production system also pays for the services wrapped around the model, the data layer underneath it, the operational overhead above it, and the people beside it.
| Component | Appears on a model pricing page? | Scales with |
|---|---|---|
| LLM inference | Yes | Tokens |
| Speech, telephony, image APIs | No — separate vendors | Minutes, calls, images |
| Vector database and retrieval storage | No | Index size, throughput |
| Evals, tracing, observability | No | A fraction of model spend |
| Retries and fallbacks | No | A fraction of model spend, higher in agentic loops |
| Fixed infrastructure | No | Provisioned peak, not usage |
Why LLM inference lands at 15–35%
The pattern repeats across enough published production breakdowns to be treated as a rule of thumb: the model is a minority of the operating cost. That is not because inference is cheap in absolute terms — for a pure text product it can be the largest single line — but because everything else arrives once the system is real: speech services on a voice product, a vector store on a retrieval product, retry traffic in an agentic loop, and a human somewhere in the escalation path.
This calculator exists to make that structure visible before you commit to it, rather than after the first invoice arrives with four unfamiliar lines on it.
What the overhead percentages actually represent
Evals and observability are deliberately expressed as a percentage of model spend, because that is how they scale — more traffic means more logged traces, more evaluation runs and more storage. Retries are expressed the same way for the same reason: the cost of a failed call is a call.
The important asymmetry is that retries do not respond to the optimisations that shrink everything else. Shortening prompts and downgrading models reduce the base, but a retry-heavy agentic workflow keeps its multiplier. If your retry rate is high, fix the retry rate; no amount of prompt engineering will help.
Where teams get this wrong
- Quoting the model bill as the project cost. It is usually the smallest number anyone puts on a slide.
- Modelling the average and provisioning for it. Reserved capacity, concurrency limits and per-seat fees are priced at peak.
- Assuming the overheads scale down with the model bill. Evals and retries track call count, not token price.
- Leaving human review out entirely. For regulated or high-stakes outputs it is often the largest single line, and it never appears on a vendor invoice.
- Comparing a self-hosted design against a hosted one on tokens only. Self-hosting trades a variable cost for a fixed one, and the fixed one is provisioned at peak.
How to use this number
Replace the defaults with rates from your own contracts, then look at the proportions rather than the total. If the model is more than half of your bill, you have either a very focused text product or a missing line. If the surrounding machinery is more than eighty percent, the cheapest improvement available to you is almost never a cheaper model — it is a better retrieval design, a lower retry rate, or a shorter prompt.
What a healthy split looks like
There is no single correct shape, but the wrong shapes are recognisable. A text-only assistant with a short prompt and modest traffic should see the model as its dominant line; if it does not, there is probably a fixed cost being spread across far too little volume.
By contrast, when a voice product or a retrieval-heavy system shows the model as the largest single line, that is worth checking. Either the other meters are not being counted, or the workload really is model-dominated — which happens when prompts are enormous. And if fixed infrastructure is the biggest line, you have a scaling problem rather than a cost problem: the cure is more volume, not a cheaper model.