Cost governance

Set limits on what your assistants may spend on AI models, so a bill never surprises you. You can stop a single conversation from costing more than a dollar, cap an experimental agent at $5 a day, keep the whole organization under $500 a month, and check almyty’s own cost figures against your model vendor’s invoice.

Spend goes wrong in two ways: one run gets stuck in a loop and burns money fast, or many well-behaved runs quietly add up. almyty has a separate limit for each.

Limit a single run

Every run has a cost cap. If you set nothing, a run may spend up to $1 before it is stopped. Change it per agent under Run limits on the agent’s edit page; see Run limits for every limit and how they combine.

The cap only knows about its own run. To limit what many runs spend together, use a spend budget.

Spend budgets

A spend budget is a ceiling across runs for a day or a month: “the organization spends at most $500 a month”, “this experimental agent spends at most $5 a day”.

Add one

1. Open Analytics → Budgets and click New budget.
2. Under Applies to, choose The whole organization or A single agent (then pick the agent).
3. Enter the Limit (USD) and pick the Period: Per day or Per month. Days and months follow UTC.
4. Under When the limit is reached, choose:
   • Warn only: runs carry on; almyty records an alert and emails the organization’s owners and admins, once per period.
   • Block new runs: once the limit is reached, new runs are refused (scheduled agents, chat replies and API calls all stop) until the period rolls over.
5. Warn at (% of limit) sends an early alert, 80% unless you change it. Runs are never blocked at that point.
6. Click Create budget.

Turn off Enforce this budget to keep a budget on record without it affecting any run. Only admins and owners can add or change budgets.

The Budgets tab shows how much of each budget has been spent this period, and Recent budget alerts lists each warning. A budget sends at most one early alert and one limit alert per period, however many runs cross it.

What counts. Budgets add up what agent runs cost, including runs that failed or were cancelled, because the money was still spent. Chatting with an agent outside a run is not counted.

The Cost tab

Analytics → Cost shows total spend for today or this month, spend per day and the agents that spent most. Below that is the provider comparison.

Compare with your model vendor’s figures

almyty’s costs are estimates, worked out from how much text was sent and received and the vendor’s price list. Your vendor’s own usage report is the real record. The comparison reads that report and shows the difference, so you notice when prices changed, when someone else is using the same key, or when an estimate is off.

It works for OpenAI and Anthropic. Other vendors are listed as not supported.

Both vendors only share usage figures with an admin key, which is a different key from the one used to run models. To add it, open the vendor’s connection on Models, go to its Settings tab, open Advanced and fill Usage key: pick an admin key you keep on Credentials or paste one. Then click Sync now on the Cost tab. Without a usage key, almyty tries the normal key, which the vendor usually refuses; the refusal shows in the result.

For developers

Per-run limits are the limits set on a run: maxCostCents, maxSteps, maxTokens, maxDurationMs, maxToolCalls, maxRecursionDepth. Defaults when nothing is passed: maxSteps: 50, maxDurationMs: 900000 (15 minutes), maxCostCents: 100 ($1), maxToolCalls: 100, maxTokens: 500000. Sub-agent runs get tighter defaults (maxSteps: 20, maxCostCents: 50). These are defaults, not the ceiling: the operator’s environment sets a ceiling nothing inside the product can exceed (RUN_LIMIT_MAX_DURATION_MS is 3600000, one hour, out of the box). A run whose accumulated cost crosses maxCostCents ends with status failed and error BUDGET_EXCEEDED. See Run limits.

Spend budget fields:

The budget form lives at /analytics/budgets/new.

Enforcement. Budgets are checked when a run starts, before anything is created, so a refused run leaves no run row behind. For each active budget matching the organization (and the agent, if scoped), almyty sums period-to-date agent-run spend:

• Soft threshold crossed: a soft alert is recorded and owners and admins are emailed. The run proceeds.
• Limit reached with behavior: warn_log: a hard alert is recorded and emailed. The run proceeds.
• Limit reached with behavior: reject: the run is refused with HTTP 403:

{
  "success": false,
  "error": "SPEND_BUDGET_EXCEEDED",
  "message": "Spend budget exceeded",
  "detail": {
    "budgetId": "…",
    "spentCents": 50213,
    "limitCents": 50000,
    "periodType": "month"
  }
}

Alerts are deduplicated per budget, per period, per level (soft/hard). Alert and email failures never block a run.

Reconciliation. One row per provider: estimateCents, estimateTokens, actualCents, actualTokens, deltaCents (actual minus estimate) and deltaPct. The actual* fields are null until a sync has run. A sync upserts one snapshot per provider per day, so re-syncing a window is safe. Unsupported provider types are reported by the capabilities endpoint with supported: false and a note; no network call is made for them. Admin keys: OpenAI needs an Admin key (sk-admin-...) with the api.usage.read scope (a normal sk-... key gets a 401); Anthropic needs an Admin API key (sk-ant-admin...). Over the API the key is usageApiKey in the provider’s configuration; it is stored as a credential and masked in responses.

API. Budgets, spend rollups, alerts and reconciliation live under /budgets and /provider-usage. Reads need member or above; changes need admin or owner; users in several organizations must send X-Organization-Id. See the API reference.

Related

• Analytics: where the Cost and Budgets tabs live
• Connect a provider: connecting a model vendor
• Enterprise: chargeback, spend per team with a forecast