Margin and Revenium: one records, one verifies and acts.

Revenium records what your AI spends and what the caller says it achieved. Its outcomes and quality scores are supplied by the caller. Margin grades the outcome with your eval, and acts on the result. On the pages we read, we did not find that on their side.

Last verified 2026-10-20. Public pages only, read that day. A page we did not find something on is a search result, not proof it does not exist.

Row by row

Cost per outcome

Different jobs

Revenium

Reports cost per job and per conversion. The outcome is attached by the caller: SUCCESS, FAILED or CANCELLED, plus a type such as CONVERTED.

Margin

Cost per outcome from the calls and a graded pass or fail, using the customer's own evals as the grade. The execution bill is included.

Sources: Revenium docs: agent outcomes, Margin: the Meter. Read 2026-10-20.

Value per outcome

Different jobs

Revenium

Records a monetary value per outcome and a per-job ROI. The caller supplies the value and picks its provenance label; we did not find it feeding a decision about which model runs.

Margin

You supply what a passed outcome is worth, and Margin weighs every switch against it: a cheaper model that loses more value than it saves is held. With no value supplied it reports cost per outcome and says so.

Sources: Revenium docs: agent outcomes, Revenium docs: ROI and unit economics. Read 2026-10-20.

Quality verification

Margin's claim

Revenium

A quality score exists as optional caller-supplied metadata. On 2026-10-20 we read 73 docs pages and 412 API operation names and did not find a grader, rubric, judge or held-out set.

Margin

A pass is graded by your own eval, through adapters for promptfoo, Braintrust, LangSmith, Ragas and DeepEval.

Sources: Revenium docs: SDK setup, Revenium API reference, Margin: the Parity Gate. Read 2026-10-20.

Action: promote and revert

Margin's claim

Revenium

Spend rules that block, run in the SDK, start in shadow mode. We did not find a route, parity, promote or revert surface in the docs or the API operation names we read on 2026-10-20.

Margin

Promotes a cheaper route only at measured parity and lower cost, keeps measuring after the merge, and opens the revert when parity slips.

Sources: Revenium API reference, Margin: the Governor. Read 2026-10-20.

What gets priced

Different jobs

Revenium

Tools, MCP servers, external APIs, SaaS and human review time are priced as events beside model calls.

Margin

Every model call and the run's execution bill, tied to whether the outcome passed, so the number is cost per passed outcome rather than cost per event.

Sources: Revenium docs index, Margin: the Meter. Read 2026-10-20.

Billing

Different jobs

Revenium

Sells a billing side too: customers, products, subscriptions, invoices, payment gateways. Its own meter is priced per write. Observe is free to 100K transactions a month; Growth lists no base price, and the overage figure appears as $1.50 and as $0.10 per 1,000 on the same page.

Margin

No billing product. A pilot, then a flat fee that does not rise with your bill.

Sources: Revenium pricing, Margin pricing. Read 2026-10-20.

Why teams choose Margin

  • The outcome is graded by your own evals, not reported by the caller.
  • It acts: a cheaper route ships only at measured parity, as a pull request you merge.
  • It keeps measuring after the merge and opens the revert when quality slips.
  • No model to sell and a flat fee, so it has no stake in which model wins.

Records versus verifies and acts

Revenium tells you what your AI spent and what the caller reported it achieved. Margin grades the outcome on your evals, decides whether a cheaper model holds, and ships the change as a pull request. On the pages we read on 2026-10-20, we did not find a grader or a route action on Revenium’s side.

Run your own numbers

Enter what you measured. The calculator shows only what those numbers support and names any field it still needs.

Your invoice total.
The part you could test a cheaper model on.
Needed for cost per outcome and passes lost.
Its measured price per task divided by the current model's. 0.4 means 40% of the price.
From your own evals.
From the same evals.
Optional. Only you can supply it.

This does arithmetic on the numbers you type. It measures nothing, and it does not know whether your pass rates or your price ratio are right.

Enter your numbers to see what they support.

See the loop on real spend.

The console shows the gate on open-source agents we metered ourselves, every number tagged real or simulated.