Sizes & Costs

Choose an orb size based on the work it needs to do. Smaller orbs are inexpensive and work well for focused tasks. Larger orbs can run bigger builds, more services, and parallel test workloads.

Available Sizes

Every current orb size includes 60 GB of disk.

Orb pricing for enterprise workspaces is shown in project settings.

a1.3xlarge is available only to Gigawatt subscribers and Enterprise workspace members, due to the general scarcity of compute.

Billing

Orb usage is billed by the minute. A paused orb costs nothing. Amp automatically pauses an orb when it is idle and pauses it right away when its thread is archived.

Opening a portal or sending another message wakes the orb. Billing resumes while the orb starts and handles the request, then stops when it pauses again.

When an Orb Pauses

An orb pauses when both of these are true:

• The agent has done no work for 5 minutes.
• You have not interacted with the orb for 20 minutes.

In other words, an orb pauses 5 minutes after the agent finishes or 20 minutes after your last interaction, whichever comes later. These are idle timeouts, not a limit on how long an orb can run. An orb keeps running for as long as the agent has work to do, and you are only billed while it runs.

Agent work is anything that needs the orb, e.g., running a command or waiting for a test suite. Some work starts without a message from you and still keeps the orb running:

• Results that other agents send back to the thread
• Automations and webhooks that wake the thread
• Plugins that hold a keep-alive lease with amp.system.executor.keepAlive()
• Messages that were queued while the agent was busy

A service that the agent left running, e.g., a dev server, does not keep the orb awake on its own. It pauses with the orb and is available again when the orb wakes.

Interactions are actions you take in the thread, e.g., sending a message or using the Files, Terminal, or Desktop panes. Any authorized HTTP request to one of the thread’s portals also counts, whether it comes from you, from someone you shared the portal with, or from a service calling the portal URL. Such a request wakes a paused orb and starts a new 20-minute window. The 20-minute window keeps the orb available while you review the result, so you do not wait for it to wake up after each action. A message from Puck or from an automation is agent work, not an interaction. It wakes the orb and the agent works on it, but the orb pauses 5 minutes after that work finishes instead of staying available for 20 minutes.

The thread’s actions menu shows whether the orb is running or paused and repeats these timeouts. To stop an orb before it pauses on its own, archive the thread.

Orb subscriptions and allowances are explained on the Pricing page.

Choose a Size

• Use tiny for research, small scripts, and focused edits with light checks.
• Use small for most single-package projects and routine coding tasks.
• Use medium for moderate builds, local services, and test suites.
• Use large for monorepos, several services, or CPU-heavy tests.
• Use xxlarge for unusually large builds and parallel workloads.

Start small when you are unsure. You can choose a larger size for the next thread without changing the project default.

Set the Project Default

Open Projects, select the project, and choose its default orb size. Personal projects for Megawatt subscribers use a1.small by default.

In Personal Settings → Project Defaults, the orb size applies to new personal projects and new threads with No Project. When you select an existing project, that project’s orb size takes priority.

The default applies to new threads. Changing it does not resize an orb that is already running.

Size per Orb

Select a size in the new-thread dialog on the web when one task needs more or less compute than usual.

From the CLI, pass --orb-size with amp -ox:

amp -ox "Run the full browser test matrix" --orb-size a1.large

When Amp delegates work to another agent, include the size in your request: