Concepts

Usage and billing

Rush meters the compute your runs use while they are awake. Every session records its tokens and cost, and limits answer with a typed code.

View .md

What a run costs comes from two places. Its model calls run on your own Claude or ChatGPT subscription or API key by default, and your provider bills those. Its compute, the sandbox the agent works in, is what Rush meters against your plan. Pricing lists the plans and what each includes.

Billing is per person, not per org. Runs you start, and runs started with an API key you created, count against your own plan whichever org they run in.

04A fifth waits: 429

ONE RUNpausedcomputestopped: $0 while parkedONE KEY · max_concurrent: 4run 1run 2run 3run 4run 5 → 429wait for one to end, or cancel one
Scene. Compute is billed while a sandbox works and stops while it is parked; a key runs at most max_concurrent sessions at once (4 by default).

What is metered

  • Compute. Measured from the sandbox's vCPU-seconds and memory GiB-seconds while it is awake. A plan includes agent-hours for the month and a smaller allowance for the week, and neither carries over.
  • Parked time is free. When a turn ends the sandbox parks and its meter stops. A follow-up starts it again. A session you keep for days between turns costs nothing while it waits.
  • Disk is not priced.
  • Rush's models. A run with "strategy": "platform", or one with no account of yours connected, uses a one-time model trial Rush covers. Past it, such a run answers 402 TOKEN_LIMIT_REACHED. See Models.

Model calls on your own account are never counted toward Rush's model limits. The compute those runs use always is.

Where usage shows

Every session carries its own usage on GET /o/{org}/sessions/{session}: input_tokens, output_tokens, cache_read_tokens, cache_write_tokens, reasoning_tokens, total_tokens, total_cost_usd, tool_calls_count and turns_count, summed across turns. The webhook for each turn also carries total_cost_usd.

total_cost_usd is the model cost of the run. On your own account it is the cost the harness reports, which your provider bills; it is not a Rush charge.

In the console, Billing shows your plan, the compute you have used this period, and the spend limit.

Limits and their codes

When a plan does not cover a run, starting it or sending a follow-up answers 402 with a typed code, before any sandbox starts:

code Meaning
COMPUTE_LIMIT_REACHED The month's included compute is used.
WEEKLY_COMPUTE_LIMIT_REACHED The week's included compute is used.
SPEND_CAP_REACHED Usage for the period reached the spend limit you set.
PAYMENT_METHOD_REQUIRED A paid plan has no working payment method.
TOKEN_LIMIT_REACHED The one-time model trial is used. Connect your own account.

Spend limit

An org admin sets a spend limit in the console under Billing. It caps what the period can be charged, and once usage reaches it new runs answer 402 SPEND_CAP_REACHED. Clear it to remove the limit.

Concurrency

Two limits decide how many runs can be in progress at once.

  • Per API key. A key has max_concurrent, set when the key is created: 4 by default, up to 100. Starting one more answers 429. Wait for a run to end, or cancel one.
  • Per plan. Your plan also sets how many of your runs can be active at once, across the console, the CLI and your keys.

A parked sandbox holds none of your plan's slots, so sessions waiting between turns do not count against it.

A running session also queues at most 20 follow-ups from the console or the CLI; one more answers 429 QUEUE_FULL.

rush http POST /o/$ORG/api-keys -d '{"name": "ci-runner", "max_concurrent": 20}'

See create an API key.