AI model economics · pilot inquiries open

The cheapest AI model isn’t always the most economical.

Understand what your AI workloads cost—and what the evidence actually supports.

Analyze your AI usage offline. Compare reported costs, inspect successful requests and completed tasks where supported, and review the evidence behind every result. No live integrations required.

Try AIYVO Today Explore the Planned Platform

Try AIYVO Today: explore a working example using synthetic data. Review the resulting metrics, comparisons and evidence directly on this website.

$ python3 aiyvo_usage.py analyze export.csvSynthetic example

The lowest cost per AI request is not necessarily the lowest cost per successful request.

  • Cost per request
  • Cost per successful request
Synthetic example: model ranking by cost per request versus cost per successful request Illustrative synthetic data on one dollar scale. Each record in this example is one AI request. model-c1: $0.0147 per request (rank 1), $0.1474 per successful request (rank 2). model-b1: $0.0510 per request (rank 2), $0.0684 per successful request (rank 1). model-a1: $0.1743 per request (rank 3), $0.2015 per successful request (rank 3). Ranks compare the three models that have both figures. $0.00 $0.05 $0.10 $0.15 $0.20 $0.25 COST PERREQUEST COST PERSUCCESS model-a1: $0.1743 per request (rank 3) model-a1: $0.2015 per successful request (rank 3) #3 model-a1 $0.1743 $0.2015 model-a1 #3 model-b1: $0.0510 per request (rank 2) model-b1: $0.0684 per successful request (rank 1) #2 model-b1 $0.0510 $0.0684 model-b1 #1 model-c1: $0.0147 per request (rank 1) model-c1: $0.1474 per successful request (rank 2) #1 model-c1 $0.0147 $0.1474 model-c1 #2
Synthetic example: model ranking by cost per request versus cost per successful request Illustrative synthetic data on one dollar scale. Each record in this example is one AI request. model-c1: $0.0147 per request (rank 1), $0.1474 per successful request (rank 2). model-b1: $0.0510 per request (rank 2), $0.0684 per successful request (rank 1). model-a1: $0.1743 per request (rank 3), $0.2015 per successful request (rank 3). Ranks compare the three models that have both figures. COST PERREQUEST COST PERSUCCESS model-a1: $0.1743 per request (rank 3) model-a1: $0.2015 per successful request (rank 3) $0.1743 model-a1$0.2015 model-b1: $0.0510 per request (rank 2) model-b1: $0.0684 per successful request (rank 1) $0.0510 model-b1$0.0684 model-c1: $0.0147 per request (rank 1) model-c1: $0.1474 per successful request (rank 2) $0.0147 model-c1$0.1474
Illustrative synthetic data on one dollar scale · fictitious models, not customer data · output of the offline M1 analysis, not a live dashboard. Ranks compare the three models that have both figures.
Synthetic

10.0×

model-c1: what a successful request costs compared with what any request costs

$0.14744 ÷ $0.01474 = 10.0, exactly 200 requests ÷ 20 successes: nine in ten attempts fail.

Synthetic

2.2×

model-c1 compared with model-b1, cost per successful request

$0.14744 ÷ $0.06842 = 2.15, shown rounded.

An illustrative ratio between two synthetic models. It is not a savings claim, a verified performance improvement or a customer outcome.

The ranking reversal

Cheapest per request. Second per successful request.

Price per request tells you what each attempt costs, not what a result costs. Failed attempts are still paid for.

Here, only 20 of model-c1’s 200 requests succeed, so each successful request costs $0.1474. Model-b1 costs more per request, yet 149 of 200 succeed, which makes it the most economical per successful request.

cost per successful request =
  cost of all requests
  ÷ successful requests

Synthetic example. Ranks compare the three models that have both figures.
ModelPer requestPer successful request
model-c120 of 200 succeeded$0.0147#1 per request$0.1474#2 per successful request
model-b1149 of 200 succeeded$0.0510#2 per request$0.0684#1 per successful request
model-a1173 of 200 succeeded$0.1743#3 per request$0.2015#3 per successful request
Calculation details and data scope
Synthetic example, fictitious models. US dollars, rounded to four decimal places.
ModelTotal costSuccessesPer requestPer successful request
model-c1$2.948920 / 200$0.0147$0.1474
model-b1$10.1953149 / 200$0.0510$0.0684
model-a1$34.8651173 / 200$0.1743$0.2015
  • Each record in this synthetic export is one AI request.
  • A successful request is not a completed task. When a task takes several requests (retries or agent steps), cost per completed task needs one row per task, with the total cost of all its requests and the task’s final outcome.
  • A model is compared only with at least 30 records, and cost per successful request requires complete cost and outcome data, so two other models in the example are not ranked here: model-d1 has 12 records, and model-e1 has no cost on 4 of its 60 records.
  • The comparison is observational: the models served different requests, so it is not a controlled experiment.

These figures are real output of AIYVO’s offline analyzer, on synthetic data. See each step that produced them. No signup is needed to see them.

Try AIYVO Today Request Pilot Access

How it works today

A local analysis of the data you already have.

AIYVO provides an offline AI usage analyzer that helps developers examine reported costs, outcomes and supporting evidence. It runs on your machine, on an export you choose.

  1. Export

    Export usage rows you already log: model, tokens, cost, outcome.

    usage-export.csv

  2. Validate locally

    Every row is checked on your machine. Columns and values that look like prompt text, personal data or secrets are refused.

    python3 aiyvo_usage.py validate

  3. Analyze outcomes

    Cost, success rate and latency per model, with coverage stated.

    python3 aiyvo_usage.py analyze

  4. Compare models

    Cost per record next to cost per successful request, or per completed task when each row is a whole task. Thin evidence is withheld.

    report.md

offlineRuns on your machine. Nothing is uploaded, and there is no live connection to any provider.

Today

What exists now, stated plainly.

Available today · offline analysis

  • todayLocal CSV / JSON import
  • todayData validation
  • todayCost-per-success analysis
  • todayWithholding when evidence is insufficient

Not provided today

  • noLive integrations or hosted dashboards
  • noVerification against provider billing
  • noEvidence of savings
  • noProcurement execution or authorization

Vision · planned

Toward governed AI infrastructure procurement.

  • planned

    Procurement intelligence

    Choosing models and providers against explicit quality, cost and latency requirements.

  • planned

    Spend governance

    Budgets and caps for AI usage, reconciled against the outcomes they produce.

  • planned

    Authorization

    Approval of AI spend by the right people, with the evidence behind each decision kept on record.

These capabilities are planned. None of them is available today. Explore the planned platform in a roadmap preview: an interactive simulation on synthetic data.

Explore a pilot on your own workloads.

AIYVO is accepting pilot-interest inquiries from teams who want to measure one real AI workload. General developer-preview invitations are not available. Sending an inquiry creates no account, subscription or commitment, and does not guarantee a pilot. Or email jorge@aiyvo.com.

Request Pilot Access