ProductDevelopersPartners
OverviewOya UniversityOAgentsAgent IDEKnowledgeAgent BrainChatEvalsChannelsRoutinesRun HistoryMonitoringOrganizationsWidgetBilling
Docs›Billing

Billing

Transparent pricing in plain dollars.

Overview

Oya bills in plain US dollars. No tokens, no credits, no hidden multipliers. You see exactly what each action costs.

How Billing Works

Oya uses a prepaid balance model. Add funds via Stripe, and they are deducted per usage. A signup bonus is provided for new accounts so you can start building immediately.

Warning
When your balance hits $0, OAgents stop responding to chat and API calls. Webhooks and routines also stop executing. Top up to resume.

Every charge includes a fixed platform margin. Underlying costs (LLM tokens, sandbox compute, tool calls) are marked up by a 60% gross margin, so the amount you pay is cost / (1 - 0.6), or 2.5× the raw cost. The figures below are the raw pre-margin rates.

Cost Breakdown

There are three cost types on the platform:

LLM Calls

Per-token pricing based on the model used. The exact cost of each LLM call is shown in the Run Tree per step, so you can see exactly where your money goes.

Per-Model Pricing

Each usage record shows the model name used for that call. Cost varies by model: lighter models like Gemini Flash cost significantly less than larger models. The model is displayed on each billing record so you can see exactly which model generated each charge.

Tip
Switch to cheaper models for simple tasks to reduce costs. The model can be configured per OAgent in the Soul step.

Sandbox Execution

Per-second pricing based on compute (vCPU, RAM, storage). When a run's duration isn't measured, a flat default of $0.001 per run (pre-margin) applies. Either way it works out to fractions of a cent per run.

Tool Calls

A small per-call fee for tools the OAgent invokes. Built-in tools (web search, fetch URL, knowledge base search) cost $0.0001 per call, and external MCP tools cost $0.0005 per call (both pre-margin).

Cache-Aware Token Billing

LLM prompt caching is billed at reduced rates. Cached prompt tokens are a subset of the input tokens, so they are not charged the full input rate: cache reads bill at roughly 0.1× the input rate and cache writes at roughly 1.25× the input rate. Only the non-cached remainder is charged at the full input rate, so re-using a large system prompt across turns stays cheap.

Tip
Sandbox execution is extremely cheap (fractions of a cent per run). The majority of cost comes from LLM calls during the first compile. Subsequent runs are sandbox-only.

Free Credits

New accounts receive a signup bonus. You can also earn a one-time $5.00 credit by connecting your GitHub account in Settings. The credit is added to your balance automatically after connecting.

Low Balance Warning

When your balance drops below $1.00, a warning banner appears showing your current balance. Clicking it opens the Add Credits dialog. At $0.00, OAgents stop responding entirely.

A run that hits an insufficient balance fails immediately and is terminal: it is not retried or requeued, because it would just keep failing until the account tops up. When this happens, the account's funding owner is emailed so they can top up (the email is deduped so repeated blocked runs don't spam). Once you add funds, new runs resume normally.

Viewing Usage

The /billing page shows your current balance, usage statistics cards (today, this week, this month, last month), and a detailed usage history table.

Filtering & Search

The usage table supports multiple filters:

  • Type filter: Show All, LLM, Tools, Sandbox, or Credits records.
  • Text search: Search by model name or OAgent name.
  • Agent filter: Dropdown to show usage for a specific OAgent only.
  • Pagination: 50 records per page with Previous/Next navigation.
Billing page showing balance, usage records with model names, and filters
The billing page with searchable, paginated usage records.

Record Details

Each usage record shows the exact dollar amount charged. Additional details vary by type:

  • LLM records: Show prompt tokens and completion tokens.
  • Tool records: Show the tool name (web_search, fetch_url, gmail_send_email, etc.).
  • Sandbox records: Show compute duration.
Warning
Usage records update in real-time. If a chat message triggers 3 LLM calls and 2 tool calls, you'll see 5 separate usage records.

Cost Insights

The Insights tab (next to Usage on the /billing page) scans your recent usage and flags why an agent is expensive, so you don't have to read raw records to find a runaway cost. Each issue is ranked by impact and comes with a concrete fix.

The Insights tab listing cost issues per agent: context bloat, one skill driving most of the cost, and oversized prompts, each with a severity and a recommended fix
The Insights tab: issues grouped by agent, each card showing the problem, the metric behind it, and a specific fix, most impactful first.

Insights detect the patterns that quietly run up a bill:

  • Context bloat: the agent re-sends a large amount of context on every step (a high input-to-output token ratio), paying to re-process the same data over and over.
  • One skill driving cost: a single skill (often a CRM or list read) accounts for most of the agent's spend because its results are large and get carried forward.
  • Oversized prompts: an individual call carries a very large input, usually meaning an entire dataset was loaded into one prompt.
  • Expensive routine: a scheduled routine is a top cost driver and processes too much in one run.
  • Spend spike: a single day cost far more than a typical day.

Cards for agents in your account also have a Fix it with Oya Engineer button: it opens the agent's Engineer with the issue and recommended fix pre-loaded, so you can apply the change without copying anything over.

Tip
Insights are computed on demand from data already recorded, so there is nothing to turn on. Only agents that exist and are switched on are analyzed; spend from deleted or offline agents is excluded. Start here whenever a balance drops faster than expected: the top card is almost always the thing to fix first.

Adding Funds

Credit top-ups are handled via Stripe as one-time payments. The minimum amount is $1.00. Funds appear in your balance instantly after payment.

Tip
On pay as you go there are no recurring charges, only pay for what you use. Add funds whenever your balance runs low, or turn on auto-recharge so you never have to think about it.

Invoices and receipts

Every payment sends you a receipt by email with the amount, date, reference, the card used, and a link to the invoice. The same payments are listed under the Invoices tab on the Billing page: credit top-ups, automatic recharges, and agent plan charges together, newest first, each with an invoice or receipt you can open and download. Anything older than the most recent payments is a click away in the Stripe portal.

Per-Agent Plan

Everything above describes pay as you go, the default: a prepaid credit balance drawn down per run. Organizations and agencies can instead pay a flat price for each live agent, billed monthly or annually like a seat. An account is on one model or the other, never both.

What counts as a live agent

An agent uses one license when it is deployed and online. Drafts and agents you take offline are free, so you can build and test as many as you like and only pay for the ones doing work. Putting another agent live when every license is in use asks you to add one first, and you can take an agent offline to free its license instead.

Usage and your credit balance

Usage is included under fair use: there is no metering, no top-ups, and your invoice does not move with a busy week. Runs are still itemized in your usage history, so the Insights tab keeps working exactly as it does on pay as you go. Your existing credit balance is paused rather than spent while you are on the plan, and the full amount is available again if you switch back.

Changing the license count

Add or remove licenses from the Billing page at any time. Stripe adjusts the next invoice for the change: adding a license part-way through a period charges the remainder of that period, and removing one credits it back. You cannot drop below the number of agents you currently have live, so take an agent offline first.

Tip
Not sure which model fits? Pay as you go is cheaper while you are building and for anything bursty or occasional. The per-agent plan wins once agents are running in production every day and you would rather have one predictable invoice than a variable one.

Cost Optimization Tips

  • Use cheaper models (e.g., gemini-3.5-flash) for simple tasks.
  • Minimize behavior rules: each rule adds to the system prompt (more input tokens).
  • Use the Run Tree to identify expensive steps and optimize.
  • For repetitive tasks, use automation scripts (import Python) instead of LLM-driven assistants.