Cost, tokens and time

An Isolation session costs what it measured: every session records its CPU seconds, memory, storage, egress, tokens and time, and prices the resource part with a rate card into one deterministic estimate. That number is for attribution, never a charge. The organization's balance pays for exactly two things, Cloud Server uptime in whole hours and egress, and a session on a Connected Server bills nothing at all. Tokens are counted per session and paid to your provider on your own credential.

Three measures on every session

Cost. The sandbox's consumption is sampled while it runs and stored on the session as cumulative totals: CPU seconds, GB-seconds of memory, GB-hours of storage and bytes of egress. A rate card turns them into micro-dollars.

The estimate is cpu + memory + storage + egress at those rates, rounded, and recomputed from the latest totals rather than summed from deltas, so a retried sample can never inflate it. The rates are published so that the same work on any server gives the same number, which is what lets you compare a feature built on a Connected Server with one built on a Cloud Server. A Cloud Server is shared by many sessions over its life, so its real charge stays on the server; the session carries the approximation. Tokens are not priced into it.

Tokens. AI credentials are personal, and the real key never enters a sandbox: every call leaves through the credential gateway, which injects the header on the way out and keeps a per-session count of what went through. session_usage reads the same story from the other end, the harness transcripts inside the session, per model, input and output. Both are observational, what the session actually consumed. The dollars, if the provider publishes a price, are shown as AI spend beside the estimate and never added to the balance. How credentials reach the gateway is in AI providers.

Time. A session records when it launched and when it closed. Elapsed time is what the session screen shows first, and a paused session's clock stops with its meter, because a frozen sandbox consumes nothing; see pause, resume, save and finish.

Where you see it

On the session screen, the session row carries "cost so far", or "free · local" when the session runs on a Connected Server. Its consumption panel breaks the estimate into the four resources, next to the token count and the elapsed time. Finished sessions keep their line in the history with the same figures.

From a chat, /status answers with the organization's servers, which one new sessions launch on, every session of yours that is running with its state and start time, and the balance.

/status
Servers: The Mac (up) · eu-1 (asleep)
New sessions launch on The Mac.
Running: login fix (ready) · search index (paused)
Balance: $42.10.

From MCP or the CLI, session_usage takes a session id and returns its token usage per model; session_get returns its state, agents and views; sessions_list returns everything of yours that is open. The cost of a session is a member's question, not the agent's.

The Stats page of the organization is where the numbers add up: AI spend and the resource estimate for the whole organization, then the same by member and by workspace, with session counts and tokens. That is the answer to "what did this feature cost?", per person and per workspace over a time window, and it has its own page: metrics per member and team.

What is actually billed

Only Cloud compute. A Cloud session's machine is billed per second, at the fixed price of the size its server was created with, from launch to the moment the session pauses or finishes; a dedicated Cloud VM is billed per hour while it is up. A paused Cloud session costs nothing. When the balance runs out, running Cloud sessions are saved and paused, and resume on the next top-up. Members never enter a card; the balance is the organization's, and the Owner funds it. The full model is in Cloud Servers and billing.

Everything a session measured stays on its row after it finishes, so a session that ran on a server you no longer have still says what it cost, in tokens, in resources and in minutes.