Metrics per member and team

To answer "what did this feature cost", every session measures what it consumed, prices it with one rate card, and the result rolls up per workspace, per member and per time window for the whole organization. A feature is a workspace (or a session of it), a person is a member, and the week is a window, so the same numbers answer all three questions. None of this is a bill. The organization's balance is a separate thing, covered below.

What one session measures

A session records five measured totals while it runs, and its wall time from launch to finish:

• CPU seconds, memory GB-seconds and storage GB-hours of the sandbox.
• Egress, the bytes the sandbox sent out.
• AI tokens, read at the credential gateway. Every request an agent makes to a provider goes through the gateway with a placeholder, so the gateway sees one row per response: agent, provider, model, input, output, cache write, cache read. session_usage reads the same numbers from the session's own harness transcripts, per model, as a cross-check.

The measured totals are priced by the rate card into a cost estimate. The defaults are 10 micro-dollars per CPU-second (about $0.036 per CPU-hour), 2 per GB-second of memory (about $0.0072 per GB-hour), 40 per GB-hour of storage (about $0.029 per GB-month) and $0.01 per GB of egress. The estimate is deterministic: the same measurements always give the same figure, on a Connected Server or a Cloud Server alike. The per-session screen is described on cost, tokens and time.

Tokens are counted, not priced, on the invoice side. Because every AI credential is personal, the provider bills the member for them. The organization's view shows the token count per session so the work is attributable, and the dollar figure a member sees is what their own provider charges.

How it rolls up

The Team page aggregates every session in the organization three ways:

• Per workspace: sessions, estimate, tokens. A workspace is how one person works on one thing, so this is the closest thing to "per feature".
• Per member: sessions, open sessions, estimate, tokens. Sessions are the launcher's, so attribution is exact.
• Per time window: the same two tables for this week, this month, or a range you pick.

Admins and Owners see the whole organization. A Member sees their own history, the same numbers for their own sessions.

A worked week

Two members, three workspaces, one Cloud-sessions server on the Standard size at $0.062 an hour, and one Connected Server. Estimates use the default rate card.

Rolled up:

So the checkout feature cost about a dollar of compute and 4.5 million tokens this week, across two people and two kinds of server. That is the accounting view. Nobody was charged $1.78.

What the balance pays for

The organization's prepaid balance pays for exactly two things: the seconds a Cloud session's machine runs and the hours a dedicated Cloud VM is up. A Cloud-sessions server itself is a configuration and costs nothing while idle; each session on it gets a machine at launch and pays the server's fixed per-hour price, per second, until it pauses or finishes. In the week above, the three Cloud sessions ran for 10 hours and 35 minutes in total, so that much time at the Standard price came off the balance, about 66 cents. The Connected Server's 14 hours cost nothing.

Below a threshold the organization is warned; when the balance runs out, running Cloud sessions are paused gracefully, each saved and snapshotted, and resume on top-up. Owners and Admins add funds with /topup <amount> or from the Billing page, in whole dollars from $5 to $500. Members never enter a card. The full rules are on Cloud Servers and billing.

AI is never on the invoice, because AI providers are personal credentials. Connected Servers never bill. And every launch, save and stop behind these numbers is in the activity log, so a figure can always be traced to who did what, and when.