Cost Analysis Template

The goal is to explain how much it cost to operate the system, why, how it scales, and which lever would reduce costs without crossing the quality gate. Use real billing/usage data and retain the current applied rate; do not copy figures from a tutorial.

1. Period and Economic Unit#

Declare:

• UTC period, environment, and commit/digest;
• total requests, successful, aborted, and failed;
• main unit: cost per completed task, not just per call;
• currency and whether taxes are included/excluded;
• date and links for each official rate;
• separate free tiers/credits from actual economic cost.

2. Real Breakdown#

Fill in with data from the dashboard and invoice:

Component	Unit	Real Volume	Current Rate	Cost	Source/Evidence
routing model	input/output tokens
generation model	input/output/reasoning tokens
ingestion embeddings	tokens
query embeddings	tokens
judges/evaluation	calls/tokens
vector DB	storage/operations
API/compute	hours/vCPU/RAM
observability	events/GB
network/egress	GB
monitoring/services	month
total

Account for evaluation, retries, and tools: they are part of the product. Show credits in a separate column; if the provider gifted €20, the observed cost may be 0, but the recurring cost is not.

3. Reproducible Formulas#

For a single call:

coste_llm =
  input_uncached_tokens / 1_000_000 * tarifa_input
  + cached_input_tokens / 1_000_000 * tarifa_cached_input
  + output_tokens / 1_000_000 * tarifa_output
  + cargos_de_tools

For a multi-step task:

coste_tarea = suma(coste_llm de cada intento)
             + embeddings_query
             + búsquedas/reranking
             + tools externas
             + porción de infraestructura y observabilidad

coste_por_tarea_exitosa = coste_total_periodo / tareas_exitosas

Do not allocate shared costs by dividing only among requests if batch, ingestion, or evaluation serve a different volume. Explain the chosen driver.

4. Distribution, not just mean#

Report p50/p95/p99 for tokens, calls, and cost per task. Add:

Segment/route	tasks	success	p50/p95 tokens	p50/p95 cost	reason
cache hit
Luna/volume
Terra/balance
Sol/badge
tool error/retry
evaluation

The GPT-5.6 names reflect the August 2026 catalog. If they change, keep the real alias and snapshot from the experiment here; do not rewrite the past with the new name.

5. 1×, 10×, and 100× Projections#

Separate variable, stepped, and fixed costs:

Scenario	tasks/month	tokens/task	cache hit	peak concurrency	variable cost	stepped infra	total
observed 1×
10× growth
100× growth
stress (p95 + retries)

Do not multiply everything linearly: an instance jumps in steps, free tiers disappear, discounts/batch may kick in, and rate limits force another architecture. Include a worst-case scenario with long input and an unstable provider.

6. Levers and Experiments#

For each optimization, calculate the measured savings and loss/risks:

Lever	Hypothesis	Monthly Savings	Quality Impact	p95 Impact	Evidence	Decision
reduce context					ablation
semantic cache					false hits + hit rate
tier routing					paired eval
batch for eval/ingestion					bill/run
local model					total GPU cost

An optimization is only promoted if it maintains critical gates. Saving 30% while worsening safety abstention is not savings: it shifts cost to incidents.

7. Self-hosting break-even#

Compare:

API mensual = volumen_tokens * tarifa_efectiva + herramientas
self-host mensual = GPU + CPU/RAM + disco + egress + observabilidad
                    + horas de operación/guardia + capacidad ociosa

Calculate required utilization and throughput, replicas for high availability, and peak reserves. Include evaluation of the quantized model. A model that fits in VRAM but does not pass the gate is not an alternative.

8. Executive conclusion#

Conclude with five figures: total cost, average and p95 cost per successful task, dominant driver, 10× projection, and savings from the next approved lever. Add the greatest risk of the projection and what signal will alert you that the assumption is no longer holding.