Skip to main content

Billing & the GOPS model

Keep bills gated agent actions by reserved capacity, not by counting individual calls. This page covers the metric, the weighting, the capacity-pack purchasing model, and the operational behavior at the ceiling; for current tier rates, see the pricing page.

1. The metric: Governed Operations, GOPS, and governed operations/min​

A Governed Operation is Keep's name for one governed agent action — the unit a policy decision produces; it is telemetry for billing, never an enforcement term. Governed Operations are metered as GOPS (governed operations per second), measured as the peak sustained load over a rolling 5-minute window. The customer-facing figure is governed operations per minute (GOPS × 120) — a whole-number rate that reads more naturally than a fractional GOPS value; the underlying measurement window is unchanged. Three names for the same thing, one per audience: governed operations/min on any customer-facing surface, Governed Operation for the thing itself, GOPS in this reference and in the API.

Within your reserved lane there is no per-call charge: governed operations inside the lane are unlimited, and there is no per-operation price at any rate — Keep sells a reserved lane, not a metered unit. The Warden gate decides allow or deny (and separately flags review_required when an action routes to review); the governed operation only weights how much that decision draws from your reserved capacity, never whether it happens.

2. What happens at the ceiling​

If sustained load exceeds your reserved GOPS, the platform does not silently throttle, reject, or queue at the gate — there is no runtime enforcement mechanism sitting between the Warden decision and your reserved capacity number. What you reserve is a billing-peak commitment: the reserved ceiling determines your bill and triggers an upsell nudge on sustained over-lane load, not a hard stop on individual calls. An agent operating near its ceiling takes zero GOPS-induced latency — the meter never sits between the Warden's decision and the agent's next step. Sizing your reservation correctly (see §7) is what keeps your bill predictable — it is a capacity-planning exercise, not a rate-limit you need to defensively code around.

3. Weighting: not every action costs the same​

A single governed operation is a sum of independently-triggerable events, each carrying its own weight and bound into the Warden gate's signed receipt at decide-time — the events cannot be altered after the fact, so the weight that's billed is exactly the weight the receipt attests to. The events are additive, not mutually exclusive: a single tool call can fire some, all, or none of them, and their weights sum.

EventWeightWhat triggers itWhat it buys
Edge Filtration0 — free, never meteredClient-side filtering (Moat + Barbican)Never reaches Keep's servers
Standard1A requires_approval_from policy clause opens a gate-token — fires identically regardless of which surface triggered itWe hold and route the decision, with a signed receipt
Semantic2Rampart classification actually runs (tool calls only)A signed, verifiable AI-threat assessment your policy acts on
Brokered4The Credential Oracle actually transfers authority for dispatch (tool calls only)Secrets sequestered in an isolated broker, the tool call's protocol handled, egress reaching only registered endpoints

Because the events are additive, a single tool call's cost depends on which of them fired:

  • A reviewed, classified, and dispatched call fires all three — 1 + 2 + 4 = 7, the full bundle.
  • A denied call with no dispatch fires Standard alone — 1 (the decision is held and recorded; nothing dispatches).
  • An auto-allowed dispatch that needs no review fires Brokered alone — 4.

A plain RLS-enforced action with no review clause emits nothing: RLS compiles to a Postgres predicate at marginal cost ≈$0, and metering it would mean metering all database I/O — the wrong posture entirely.

The weights are stable service-class coefficients derived from a documented functional decomposition of the trigger column above — not customer risk, liability, or a cost markup. Never read a weight as a surcharge without what it buys attached, since the number alone invites that reading.

Two things worth being precise about:

  • Brokered names what Keep does, not where the action goes: a no-credential outbound call brokers nothing and meters Standard alone, because none of the brokering work the weight recovers happened.
  • Irreversibility. An irreversible action doesn't necessarily require a human, and a human-approved action isn't necessarily irreversible — consequence and approval mechanism are different dimensions. IRREVERSIBLE_ACTION_GATE remains an enforcement primitive; it is not a pricing event and carries no weight of its own.

Semantic's weight of 2 is a stable service-class coefficient, not a GPU-cost pass-through: your SDK invokes Rampart directly and attaches the signed receipt, so Keep bears no compute cost at gate time for a classification that already happened. The weight recovers the value of the signed, verifiable assessment itself. And the events only weight the meter, they never gate: enforcement is grants, RLS, and the Warden gate's allow/deny decision. A governedOpClass can be escalated by policy (marking a tool more sensitive than its default), and every escalation is audit-recorded; it can never be silently downgraded below its declared or default floor to dodge the meter.

4. Capacity packs: the purchasing unit​

Keep's reserved lane is bought as a capacity pack — one uniform unit combining every dimension governed traffic draws on, so a customer never buys throughput, inference, and review capacity as separate order lines.

One capacity pack contains:

  • 1 GOPS (120 governed operations/min) — the only expansion lever; packs are sold in single units
  • up to N tokens/sec of reserved agent-reasoning inference — see §6
  • ARC (Active Review Cases) — a fixed allotment per pack; see §6

Every entitlement scales as packs × one fixed row — there is no per-tier exception table, and no tier carries an uplift or a shortfall. Platform tiers bundle a starting pack count; Traverse is the name for packs purchased beyond that bundle, on top of any tier. It is an add-on, never a standalone subscription, and it is not a rung above your tier. Exact pack counts and current pricing live on the pricing page.

There is no per-operation price, no per-event overage, and no cumulative monthly governed-operation allowance to spend down — a capacity pack reserves throughput, and reserved capacity is purchased in advance, never metered after the fact.

5. Verdict-neutral metering​

A governed operation is billed at the sum of its governedOpClass event weights regardless of the gate's verdict — allow and deny are one meter, not two. A denied call costs exactly what an approved call costs for the same set of events fired.

This is deliberate: metering a deny for less than an allow would give the platform a financial reason not to deny. Keep's billing has no stake in which way the gate rules — allow and deny draw from your reserved capacity identically.

6. Agent reasoning and review capacity​

Two further resources draw from the same capacity pack, each with different technical properties than the GOPS lane:

Agent reasoning (TPS). Reserved inference throughput for the agent's own model calls — tokens per second, independent of how many (or how few) of those reasoning steps result in a gated action. GPU capacity is scarce in a way CPU is not, so this lane is admission-controlled with no burst: a guarantee up to the reserved ceiling, never an elastic overage. Figures are always published as "up to N tokens/sec," since throughput varies by model class. Inference on Kilimanjaro's own infrastructure draws from the pack; inference on partner-served models is metered per token against a separate balance (Usage Credits), not from the pack.

Active Review Case (ARC). When a decision routes to review — review_required — the pending case holds one Active Review Case from your reserved allotment until it resolves: approval, denial, expiry, cancellation, withdrawal, or terminal failure all release it. Review is not necessarily human — the reviewing party can be a person, another agent, or a policy-defined combination; ARC counts the pending case regardless of who or what resolves it. A reviewed action meters its underlying Standard event once, when the gate opens; the review does not meter again when the verdict returns. Exhausting your ARC allotment fails closed by default — this is a durable safety review, not a queue that should silently back up.

Neither TPS nor ARC is independently purchasable: both scale with the GOPS in your capacity pack, and neither is ever billed as event-level overage.

This is a standalone meter — it doesn't interact with any other Kilimanjaro module's billing. See Keep overview §7 for the explicit boundary.

7. Sizing your reservation​

Because billing is peak-based, size your reservation to your peak sustained load over any 5-minute window, not your average — a bursty workload with a high average-but-low-peak profile is cheap; a workload with a short, sharp spike still needs a reservation that covers the spike, even if its average is low. The pricing calculator on the pricing page turns your expected monthly volume per event into a recommended tier plus capacity packs.

8. See also​

  • API reference — the governedOpClass field on the Warden gate response.
  • Architecture — how a gate decision is made and executed.
  • Audit & receipts — the signed receipt every billed decision produces.
  • Pricing page — current tier rates, capacity-pack pricing, and the sizing calculator.