Skip to content

AI and cloud costs with an owner

Cost Management · FinOps for AI and cloud

Every model call and every hour of Compute goes into the month's usage by model, key and project, in dollars and in reais, with budgets that warn or block. Alongside sit the cloud bill by service and team, Kubernetes cost by namespace, cost per business unit, idle AI tool seats and savings recommendations with evidence.

The problem

What this product solves

An AI API key on a corporate card is spend with no owner: nobody knows which project used it or why.

By the time the provider invoice arrives, the money is spent and nothing can change.

The same goes for cloud and AI coding tools: the bill grows by service and by seat, and nobody ties the amount to a team.

updated

Who uses it, and for what

Typical profiles, no customer names. Each use carries the status of the feature behind it.

  • Finance at a software companySee what AI cost per project and per key, and how much of the cloud bill belongs to each team.AI Costs
  • Engineering leadershipKnow how many Copilot, Cursor and Claude Code seats sit idle and what that costs each month.AI Tool Seats
  • Product team using ComputeKeep each application's spend within the workspace cap.Compute Costs

Features

What the product does

Self-service: add it from your account.

  • AI Costs

    AI consumption by model, key and project, in USD and BRL, with budgets, alerts and owned saved perspectives that group cost by business rules.

  • Billing

    Prepaid balance, top-ups, invoices and AI usage in BRL.

  • Compute Costs

    Metered Compute usage per deployment, with a monthly cap.

  • AI Tool Seats

    Monthly seat cost for Copilot, Cursor and Claude Code, idle seats (no use in 30 days) with waste in USD and BRL, and Cursor on-demand spend; the screen lives in Engineering Intelligence.

  • Cloud Costs

    AWS and GCP billing, OpenAI and Anthropic costs through each cost API, or FOCUS CSV, normalized per service, team and catalog service, with categories and shared-cost split, budgets, anomalies, rightsizing suggestions and cost questions by team and period.

  • Optimization

    AI savings recommendations (oversized model, bloated prompts) with evidence, estimates, human approval and realized savings after approval. Cache is not measured; cloud savings are covered by cloud costs and allocation.

  • Cost allocation

    Kubernetes cost per cluster and namespace with idle cost, cost per business unit, asset rules as code with recommended fixes, commitment and cluster-scaling recommendations. Recommendations only: nothing is bought or changed in the cloud.

  • AutoStopping

    Stops idle Deployments and wakes them on demand through the delivery agent installed in the customer cluster: a rule (idle minutes and CPU threshold) only applies after another person approves it by hash, freezes and denying policies win over approval, and AIUBY stores no cluster credentials. Waking is a request in the app, not the first incoming request; Kubernetes delivery Deployments only.

AI Costs: in detail

  • Prepaid balance in BRL, topped up via Pix
  • Price that tracks the official dollar rate (PTAX from the Central Bank of Brazil), with every adjustment in the audit log
  • BRL amounts in chat, agents, usage, budgets and the audit log

What it does not do

  • The platform does not issue invoices (nota fiscal) automatically.
  • Models run on the providers' servers; processing may happen outside Brazil.

Exchange rate and limits

From US dollars to Brazilian reais, with no surprises

The path of every model call, from the provider price to the charge against your balance. The numbers below are rules in the code; the list price per dollar is on the pricing page, and today's rate shows in your account.

  1. Provider cost

    Every call through the platform records the model, tokens and the cost in US dollars from the provider price list.

  2. Organization price per dollar

    The dollar cost is multiplied by your organization's price in reais per dollar, and any fraction of a cent goes up to the next cent. That price starts from PTAX (the Central Bank of Brazil reference rate) and adds exchange costs, taxes and AIUBY's margin; it is set per organization and shows in your account. The list price is on the pricing page.

  3. PTAX in automatic mode

    The PTAX selling rate from the Central Bank of Brazil is checked every 30 minutes. The price changes only when the variation exceeds the policy minimum, which starts at 2%, and every adjustment goes into the audit log. A contract with a fixed price does not track PTAX.

  4. Organization cap

    With self-service, the organization starts with a monthly usage cap of US$ 100. The owner can adjust it up to that amount; above it, only through the team. At the cap, the call is refused.

  5. Budget per workspace

    Each workspace has a monthly budget that blocks at the limit; with no budget set, the call is refused. Extra rules per person, key, model or project either block or only warn.

  6. Alert

    At 80% and 100% of each budget, the warning shows on screen and in the audit log. The email to the organization owners goes out once per limit per month, at the first level reached.

  7. Prepaid balance

    The amount in reais comes out of the balance. A call in progress reserves the maximum expected cost; with no balance, the call is refused until the next top-up.

How it works

From start to audit record

  1. Call

    Every model call through the platform records model, tokens and cost, per key and per project.

  2. Budget

    An organization cap and per-workspace budgets that block at the limit, with two-level warnings.

  3. Balance

    The cost is drawn from the prepaid balance. With no balance, the call is refused.

  4. Compute

    Each deployment is metered by time in use and stops when the balance or the workspace cap runs out.

  5. Cloud

    AWS, GCP, OpenAI and Anthropic billing is read, normalized by service, team and category, and checked against budgets and anomalies.

  6. Allocation

    Cluster cost is split by namespace, with idle capacity in view, and the month's cost is divided by the volume of customers, orders or transactions.

  7. Recommendation

    Rules flag models larger than needed and bloated prompts. A person accepts or dismisses each one.

Features

Feature by feature

What each one does, how it works, an example from the documentation and what it does not do. Each badge comes from the catalog.

AI Costs

Every model call has an owner, in USD and BRL

Every model call that goes through the platform Gateway (Assistant, agents, flows, indexing and API) records model, tokens and cost. Monthly usage shows by model, key, person and project, in USD from the provider price list and in BRL at the organization's rate.

Limits sit on top of that record: the organization's monthly cap, each workspace's budget and, with Guard, rules per person, key, model or project that either block or only warn.

How it works

  • The x-aiuby-project header tags a call with a project, for budgets and reporting.
  • The organization cap applies to the sum of all workspaces; each workspace budget can go up to that cap.
  • At the limit, the call is refused with 429 budget_exceeded (workspace or rule) or 429 org_budget_exceeded (organization).
  • Warnings at 80% and 100% of each budget, on screen and in the audit log; the email to owners goes out once per limit per month.
  • Lowering the cap also lowers any workspace that would end up above it; every budget change is recorded in the audit log, with the before and after values.
  • Saved perspectives with an owner, private or shared: filters, dimension and business rules (exact value or prefix) that group AI cost into buckets such as product or customer. Whoever opens one sees only what their own access allows.
Split cost by project Every call from this client counts toward the atendimento-web project's usage.
const client = new OpenAI({
  apiKey: process.env.AIUBY_TOKEN,
  baseURL: 'https://aiuby.com/api/gateway/v1',
  defaultHeaders: { 'x-aiuby-project': 'atendimento-web' },
})

What it does not do

  • Calls made directly to the provider, with your own key outside AIUBY, do not show per call; total spend on OpenAI and Anthropic comes in through the cloud cost connector.
  • Conversation text is not part of the usage record: only model, tokens, cost, workspace and key.
  • A budget cannot be removed, only changed; zero blocks usage.
  • Above the self-service cap, the amount can only be changed by the AIUBY team.

Billing

Prepaid balance in BRL: spending stops when the balance runs out

AI usage and subscriptions are drawn from a prepaid balance in BRL, topped up via Pix by the organization owner. Each call deducts its BRL amount from the balance; when the balance runs out, calls stop, and nothing is charged after the fact.

The screen shows balance, top-ups, statement and invoices. Providers charge in USD: the BRL amount of a call is its provider cost converted at the organization's rate, with any fraction of a cent going up to the next cent.

How it works

  • Top-up via a Pix code the owner generates on the Billing screen.
  • In automatic mode, the BRL rate follows the Central Bank of Brazil PTAX rate and is only adjusted when the change exceeds the policy minimum.
  • Every adjustment to the BRL rate is recorded in the audit log.
  • A call in progress reserves its maximum expected cost; with no balance, the call is refused until the next top-up.
  • Each billing situation has its own API error code, so your system can handle it.
  • If the organization is closed, the AIUBY team refunds the remaining balance via Pix.
API responses by billing situation The balance is a limit separate from the budget: with no balance, the response is 402 even if budget remains.
402 billing_account_required   Organization without a billing account
402 insufficient_credit        Prepaid balance at zero
402 subscription_suspended     Subscription unpaid past the due date
429 billing_limit_reached      Monthly billing limit reached

What it does not do

  • There is no automatic top-up, no top-up through an API key and no boleto for prepaid.
  • Only the organization owner can top up.
  • It does not issue tax invoices (nota fiscal) automatically.
  • Fixed-price contracts do not follow PTAX.

Compute Costs

Applications stay up only while they fit the balance and the cap

Compute usage is metered per deployment and drawn from the same prepaid balance. On CPU, billing is for time in use: the application stays up until you stop it, the balance runs out or the workspace cap is reached; a stopped or idle application costs nothing.

With GPU, billing is per minute the instance is on: every started minute counts, with a 1-minute minimum, including startup and idle time before the instance shuts down.

How it works

  • Monthly Compute cap per workspace, set before the first paid deployment.
  • A statement with the metered and billed periods for each application.
  • With GPU, the balance must cover the deployment minimum plus 20 minutes of the card, and 20 minutes for every other GPU application that is running.
  • With a GPU in use, if the balance or the workspace cap does not cover the next 20 minutes of the card, the application is stopped.
  • Each replica counts as one more application in the minimum balance, and the cap applies to the total.
  • Balance, cap, monthly spend and statement are also available through the CLI.
Through the CLI Both commands read the workspace of the key in use.
aiuby billing   # Compute balance, cap and monthly spend in the workspace, plus sizes with prices
aiuby usage     # Compute statement: metered and billed periods

What it does not do

  • The Compute cap is separate from the AI budget: one does not cover the other.
  • The hourly price of each size is on the pricing page, at the reference PTAX rate; the daily value, which follows the exchange rate, appears on the Compute billing screen.
  • With GPU, a started minute is charged in full, and idle time before the instance shuts down also counts.

AI Tool Seats

What AI seats cost and which ones sit idle

Connects the admin APIs of GitHub Copilot, Claude Code and Cursor and shows, per team, the monthly seat cost, idle seats and waste in USD and BRL, plus Cursor on-demand spend. The screen lives in Engineering Intelligence, next to the PRs merged in the same window.

Nothing is installed on machines, nothing is scraped from screens and no prompt or code is read: only seats and aggregated daily counts come in.

How it works

  • An idle seat is a paid seat with no use in 30 days; a seat assigned less than 30 days ago with no use is "new", not idle.
  • Monthly waste = idle seats × seat cost, in USD and in BRL at the latest stored PTAX rate.
  • The link between the tool account and the person is suggested by email and only counts after an owner or admin confirms it.
  • Default view by team; for non-admins, teams with fewer than 3 people are grouped under "Other teams".
  • Per-person view only for owners and admins, in alphabetical order, with no ranking, and recorded in the audit log.
  • Each tool's credential is stored encrypted; removing the connector deletes its credential, seats and usage.
Through the API Returns adoption and spend per team; the API never returns per-person data.
GET /api/v1/engineering/ai-efficiency?days=30
Authorization: Bearer <key with engineering:read>

What it does not do

  • Seat cost does not come from any API: the admin enters the contract value when connecting.
  • Copilot provides no tokens or usage cost, only seats and activity; Cursor provides no daily cost, only on-demand spend for the cycle.
  • Seats on claude.ai Team or Enterprise plans and usage signed in with a subscription are not covered.
  • Merged PRs next to usage are correlation, not a measure of individual productivity.

Cloud Costs

Your cloud bill by service and team, in BRL

Reads billing from AWS (Cost Explorer), GCP (billing export to BigQuery) or a CSV in the FOCUS format, normalizes everything into the same shape and converts it to BRL at the PTAX rate. Cost shows by provider service, account, region, team and engineering catalog service.

Cloud access is read-only: nothing is created, stopped or purchased in your account.

How it works

  • Connectors for AWS, GCP, OpenAI and Anthropic (organization cost API, with the admin key), with encrypted credentials and periodic collection; FOCUS CSV for everything else.
  • Cost distributed by the team and service labels, and categories with compound rules (service, account, region and label) with shared-cost split, proportional or by fixed percentage.
  • Monthly budget by total, service, account, team or catalog service, with warnings at 80% and 100%.
  • Daily anomaly detection per service against the median and deviation of the previous 30 days; only increases become alerts, and only with at least 14 days of baseline.
  • Downsizing suggestions, with estimated savings, when the CPU and memory peak sent by the customer stays low for seven days.
  • Each service's cost goes into the engineering graph: questions about cost since the latest deployment, by team and over periods of up to 92 days, with the evidence for each line item, also through the MCP server.

Example use case

A platform team connects the GCP billing export with a read-only credential and sets a monthly budget for the payments team. The following week, the daily cost of a database service rises well above the monthly median: the anomaly shows up with the service, the day and the account numbers, and whoever sees the dashboard goes from there to the catalog service and to the release that shipped the day before.

What it does not do

  • It does not create, stop or purchase cloud resources: access is read-only.
  • Direct connectors only for AWS, GCP, OpenAI and Anthropic; other providers come in through FOCUS CSV, up to 5 MB and 50,000 rows per file.
  • CPU and memory usage for rightsizing is sent by the customer; the platform does not read the provider's monitoring.
  • Cloud recommendations are text only: no action is applied. Cost drops do not trigger alerts.

Optimization

AI savings with evidence, an estimate and a human decision

AI savings recommendations calculated by rule, on the usage the Gateway records over 30 days: a model larger than the workload needs, and bloated prompts. No AI decides and nothing is switched automatically: each recommendation keeps the evidence and the human decision.

After approval, the screen compares the target's cost per call over the previous 30 days with the cost afterward and shows the realized savings.

How it works

  • Oversized model: at least 50 calls with short inputs and outputs, and a model allowed in the workspace priced at least 30% lower for the same workload.
  • Bloated prompt: at least 50 calls where the largest inputs exceed three times the median.
  • Each recommendation includes the window, query, filters, numbers and thresholds that produced it.
  • Accepting or dismissing belongs to the organization owner, and the decision is final; a decided topic does not come back for 30 days.
  • Generating, accepting and dismissing are recorded in the audit log.
  • Before switching models, the recommendation calls for an eval with real inputs from the target in AI Engineering.
Decide on a recommendation UI routes: session only, and the decision requires the organization owner.
POST /api/platform/finops/optimization          # generates the organization's recommendations
PATCH /api/platform/finops/optimization/{id}
{ "decision": "accept" | "dismiss", "note"?: string }

What it does not do

  • Cache is not measured: the Gateway does not record cached tokens per call, and the screen says so per provider.
  • It covers AI only. Cloud rightsizing and idle resources belong to cloud costs and cost allocation.
  • It does not switch models or edit prompts: accepting records the decision, and whoever owns the target makes the change.
  • API keys cannot access recommendations, and there are at most 6 generations per hour per organization.

Cost allocation

Know who spends the cluster and what each customer costs

Cost allocation completes cloud costs with what the bill does not show: how much each Kubernetes namespace spends, how much it costs to serve a customer or process a transaction, and which resources break your team's rules. Commitment and cluster-scaling recommendations point to where to save, with the math done.

How it works

  • Kubernetes cost per cluster and namespace: the collector on your network sends, daily, the cost, capacity and usage of each namespace. The split uses half CPU and half memory, and the remainder shows as idle.
  • Unit cost: you send the daily volume of a business metric (customers, orders, transactions), in total or per service, and the platform divides the cloud cost for the same period by it.
  • Asset rules as code: JSON conditions (service, region, missing label, no owner, minimum cost) evaluated on the collected resources, each violation with the recommended fix.
  • Commitments: for each service with stable spend, the daily level always in use becomes the suggested commitment, with savings estimated from the discount you provide.
  • Cluster scaling: peak usage against capacity, the node count suggested to sit at 70% at peak, and the share of spot nodes.

Example use case

The platform team turns on the collector in both production clusters and starts sending the daily number of transactions. The same screen shows the payments namespace as the biggest spender, the month's cost per transaction and a cluster that holds 40% usage at peak with two more nodes than it needs. The decision to reduce nodes or buy a savings plan stays with the team, in the provider console.

What it does not do

  • It does not buy reservations or savings plans and does not change nodes, clusters or resources in the cloud: it only recommends.
  • Kubernetes usage depends on the customer collector; the platform does not enter the cluster to measure.
  • Unit cost only exists on days when the metric volume was sent.

AutoStopping

Environments stop when nobody uses them and come back on request

AutoStopping shuts down an idle Deployment and starts it again when someone asks, without AIUBY entering the cluster: the delivery agent installed by the customer does the work, with a signed rule that only takes effect after another person approves it.

How it works

  • Rule per Deployment: consecutive idle minutes and the CPU threshold (summed across pods, from metrics-server) below which the service counts as idle.
  • Approval by another admin, bound to the rule hash; freezes and denying policies win over approval.
  • Stopping sets replicas to zero and stores the original count on the Deployment itself; waking restores that count on the agent's next cycle.
  • Turning the rule off (or the agent losing the rule) restores the replicas; the intentional divergence does not show as drift.
  • State reported by the agent (running or stopped) on the Kubernetes delivery screen, and every request, approval and wake-up in the audit log.

Example use case

The team turns on AutoStopping in the staging environment: after 30 minutes without CPU above 20 millicores, the agent scales the store Deployment to zero replicas. In the morning, whoever is testing clicks wake up and, on the agent's next cycle, the three replicas come back.

What it does not do

  • Waking is a request in the UI: the first request to the stopped service does not wake it.
  • It only applies to Deployments delivered by AIUBY in the cluster, with the agent and metrics-server installed.
  • Without CPU metrics, the agent stops nothing.

Governance and security

Rules the code enforces

  • An organization cap and per-workspace budgets. Going above the initial cap requires our team.
  • Every budget change and every automatic adjustment of the price in reais is recorded in the audit trail.
  • Anyone who can see the whole organization can follow the balance. Only the owner loads it.
  • Access to the customer cloud is read-only: nothing is created, stopped or purchased. Commitments, cluster scaling and rightsizing are recommendations, and the team decides.
  • Savings recommendations are never applied automatically: only the organization owner accepts or dismisses them, and the decision is final.
  • AI tool spend is shown per team. The per-person view is limited to owner and admin, with no ranking, and is recorded.
  • Cloud and AI tool credentials are stored encrypted and never appear in audit, logs or responses.

Integrations

What it connects to

Only integrations that exist in the code. Each one carries the status of the feature that uses it.

  • Mercado Pago (Pix)Loads the prepaid balance.Billing
  • Central Bank of Brazil PTAXReference rate for the price in reais in automatic mode.Billing
  • AWS Cost ExplorerReads AWS billing, read-only.Cloud Costs
  • Google Cloud (BigQuery)Reads the GCP billing export.Cloud Costs
  • CSV in FOCUS formatImports billing from any provider.Cloud Costs
  • OpenAI and AnthropicOrganization cost through each provider's cost API, with the admin key stored encrypted.Cloud Costs
  • GitHub CopilotSeats and activity through the admin API.AI Tool Seats
  • Claude CodeUsers and usage report through the Anthropic admin API.AI Tool Seats
  • CursorMembers, daily usage and on-demand spend through the team Admin API.AI Tool Seats
  • aiuby CLICompute balance, cap and statement from the terminal.API, CLI & keys

FAQ

Straight answers

What can I use today?

Available: AI Costs, Billing, Compute Costs, AI Tool Seats, Cloud Costs, Optimization, Cost allocation, AutoStopping. Self-service: add it from your account.

Does it import my AWS or Google Cloud bill?

Available: Cloud Costs, Cost allocation. Yes, through Cloud Costs: it reads AWS Cost Explorer, the GCP billing export in BigQuery, the OpenAI and Anthropic cost APIs or a CSV in FOCUS format, with read-only access, and shows cost by service, account, team and category, with shared cost allocation, budgets and anomalies. Kubernetes cost by namespace comes from the collector on your network, through cost allocation. It never creates, stops or buys anything in the cloud.

What do I need to subscribe to for AI Costs?

Available: AI Costs, Cloud Costs, Optimization, AI Tool Seats. AI Costs, Cloud Costs and Optimization depend on the Guard contract, the same one as AI & API Security, which has a monthly fee. Billing and Compute Costs come with Models and Compute usage, with no monthly fee. AI Tool Seats depends on Cortex. Pricing is on the Pricing page.

Can I pay by card?

No. The balance is prepaid, and pay-per-use billed to a card is not offered.

Does the platform issue tax invoices (nota fiscal) automatically?

No.

Where is usage data stored?

Available: AI Costs. Each call becomes a record of model, tokens, cost, workspace and the key used. Conversation text is not part of that record. Calls made directly to the provider, with your own key outside AIUBY, do not appear. AIUBY hosts the platform and its database. Organizations share the database, and each one only sees its own rows, enforced by row-level access rules. The daily backup is encrypted and stored on AWS in the United States.

Can I see how much each developer spends on Copilot or Cursor?

Available: AI Tool Seats. Per team, yes: AI Tool Seats shows the monthly seat cost, seats with no use in 30 days and the waste in dollars and in reais. The per-person view exists only for owner and admin, in alphabetical order, with no ranking, and every lookup is recorded in the audit trail. No prompt or code is read, and nothing is installed on machines.

Do savings recommendations swap the model on their own?

Available: Optimization, Cloud Costs, Cost allocation. No. Optimization computes by rule, with no AI deciding, and shows the evidence and the estimate. The organization owner accepts or dismisses, and the swap is made by whoever owns the target, after evaluating the candidate model. Caching is not measured. In the cloud, rightsizing, commitment and cluster scaling suggestions also only recommend: nothing changes in your account.

Works together with

  • AI EngineeringEvery agent, flow and model call goes into the month's usage, and a recommended model swap is evaluated there first.
  • Software DeliveryCompute applications are billed from the same prepaid balance.
  • Engineering IntelligenceAI tool seats appear next to each team's merged PRs.

Cost Management

Self-service: add it from your account.