Inference Cost Index
What it costs, in reais, to actually run AI: a fixed basket of tasks, repriced every month across open and closed models, served here and abroad.
- Under evaluation · no date
- Monthly · first working day
- Free, open data.
- 2027
A consumer price index for inference: one number a month that says whether building with AI in Brazil got cheaper or dearer — and why.
- Whoever pays the token bill and needs to forecast next quarter.
- Whoever chooses between an open model served in Brazil and a foreign API in dollars.
- Whoever writes about AI cost and today cites dollar list prices, without a task.
- The number
- A base-100 index, in reais, published on the first working day of the month, with monthly and annual change.
- The basket
- Twelve fixed tasks — classification, extraction, RAG, code, agent — with public prompts and volumes.
- The decomposition
- How much of the change came from list price, exchange rate, efficiency (tokens per task) and model switching.
- The data
- CSV with every task, model, provider and price, every month, to rebuild the index your way.
- 1Fix the basket
Twelve real tasks, with a mid-sized company's typical monthly volume. The basket changes once a year, with notice.
- 2Measure tokens
Every task actually runs on every model in the basket: input and output tokens measured, not estimated.
- 3Price
Provider list price × tokens × monthly average exchange rate. Brazilian providers in reais directly.
- 4Publish and decompose
The index, the decomposition and the CSV. With a paragraph from the reader: why it moved.
- 2027Go/no-go decision, after six months of Cinco tracking prices
The basket, in draft
Twelve tasks; here, five. All with public prompt and volume on the day the index exists.
- Classify
- 50,000 support tickets a month into twelve categories.
- Extract
- 10,000 invoices into structured fields.
- Answer with RAG
- 20,000 questions over a base of 2,000 documents.
- Code
- 2,000 bug-fix tasks with a test.
- Act
- 500 cases of an agent with five tool calls each.
Every week Cinco records a price per million tokens. None of those numbers says, on its own, whether building with AI in Brazil got cheaper: the exchange rate moves, models change efficiency, local providers come and go. The Inference Cost Index is the attempt to answer with one number, every month, in reais.
It is the house's most 'market intelligence' project, and the one most dependent on rigor: a fixed basket, tokens measured rather than estimated, and a decomposition separating price, exchange rate and efficiency. That is why it is still a concept, with a decision scheduled — not a promise.
- Why only a concept?
- Because actually measuring tokens on twenty models every month costs money and time. The decision comes after six months of Cinco tracking prices.
- Does quality count?
- Not in the index — cost is cost. But each task publishes a minimum quality bar to enter the basket.
Get the next Cinco
Tuesday, 7am BRT, in your inbox. Five items, with sources. No daily newsletter, no promotions, and unsubscribing is one click.