# Inference economics: reproducible cost model

Version: 2026-09-22. Author: Nick Cerutti.

**Calculated scenario, not a measured serving benchmark.** The prices and traffic below are hypothetical. No named model, hardware configuration or client outcome is implied. Actual latency, quality and utilization are unmeasured.

## Workload and assumptions

| Input | Assumption |
| --- | --- |
| Completed tasks per day | 10,000 |
| Tokens per call | 2,000 input, 400 output |
| Premium price per million tokens | USD 3 input, USD 15 output |
| Smaller model price per million tokens | USD 0.50 input, USD 2 output |
| Tasks initially routed to smaller model | 70% |
| Smaller-model calls escalated to premium | 20% |
| Other costs | Excluded: embeddings, cache storage, retries, tools, network, GPU hosting, tax and operating labor |

## Arithmetic

Baseline premium call: `(2,000 × 3 + 400 × 15) / 1,000,000 = USD 0.012`.
Baseline daily model cost: `10,000 × 0.012 = USD 120`.
Routed system: 7,000 smaller-model calls at USD 0.0018, plus 4,400 premium calls at USD 0.012. The premium count includes 3,000 direct calls and 1,400 escalations. Both calls are charged on an escalation.

Routed daily model cost: **USD 65.40**. Calculated reduction: **45.5%**, conditional on these assumptions. This is not a savings forecast.

## Reproduce

Download [reproduce-samples.mjs](https://nickcerutti.com/evidence/reproduce-samples.mjs) and run `node reproduce-samples.mjs` with Node.js 22 or later. Assertions check the arithmetic; the printed JSON separates calculated results from unmeasured properties.

## Before making a deployment decision

1. Replace the assumptions with a representative, redacted workload and current contracted prices.
2. Define task-success and safety acceptance criteria before comparing systems; count failed and retried tasks in cost per successful task.
3. Freeze dataset, model revisions, prompts, routing threshold and evaluation method. Split tuning data from held-out evaluation data.
4. Measure p50/p95/p99 end-to-end latency, escalation latency, TTFT, inter-token latency and quality at the same load.
5. For self-hosting, record GPU model/count, power, precision, engine revision, batch sizes, context distributions, cache hit rate, isolation policy and utilization.
6. Include infrastructure and operational costs. Report confidence intervals and unfavorable results; do not extrapolate synthetic replay to fleet savings.

## Engineering references and limits

[Meridian's versioned verification matrix](https://github.com/angelnicolasc/meridian/blob/f3e7cd2c127d6188c423bce18c9d4ace4454d323/README.md) distinguishes component/synthetic work from unverified real-vLLM experiments. [Stratum](https://github.com/angelnicolasc/stratum) explores dual-tier serving. Neither repository was benchmarked on a GPU for this sample.

Discuss a measured workload study: https://nickcerutti.com/services/llm-cost-sprint