Appearance
Usage, limits & billing
Usage is measured in USD across AI usage, VM compute, VM storage, and sandbox usage. Unclassified recorded costs appear under Other. Budgets control subsequent work using recorded usage; metering delays and already-started work can take usage above a limit. A connected user can optionally use a ChatGPT subscription for eligible new agent turns and memory jobs; those turns follow the displayed subscription and budget preferences rather than silently switching providers.
Capabilities
Hierarchical USD budget limits
USD limits apply at the organization, agent, room, and task levels, including ancestor-task limits. Recorded usage above an applicable limit blocks further agent work. A zero room or task limit means unlimited; a zero organization or agent limit is a cap. Raising the blocking limit allows work to resume. Room and default task budgets can be set when a Room is created.
Limits status & enforcement
A status API reports which agents, rooms, tasks, or the org have exceeded their limit, driving banners and realtime updates.
Usage reporting & windows
The workspace Usage view shows lifetime organization credit consumption against its limit. Spend over time groups charged credits by category, with AI usage expandable by purpose. VM cards separate compute and storage, with active and hibernated monthly projections. All-time room and agent totals open category breakdowns. The spend chart and VM usage use a 12-hour or 7-day window.
Organization totals count each recorded charge once. Room, agent, and task totals show metered value, which can differ from organization credits for BYOK. Rooms include owned VM compute and storage plus attributed sandbox usage; agents include owned sandbox usage plus attributed VM compute. Tasks include AI usage and attributed VM and sandbox execution costs. Overlapping work in different tasks can attribute the same resource interval to each task, so task totals need not add up to organization charges. Category breakdowns are reported as AI usage, VM compute, VM storage, sandbox usage, or Other, and resource watermarks distinguish settled charges from later execution observations.
Room usage
A room's usage page summarizes spending and its category breakdown against the room limit, and plots root-task costs by category in creation order. Each root-task total includes its subtasks. The chart includes a trailing 10-task average, opens the task from its plotted entry, and provides older/newer navigation for longer histories. Task shares use the room total and need not sum to 100%. Task and room spend rings and the agent overview usage value also open category breakdowns.
In-app budget increase
When a cap is hit, users get one-click suggestions to raise the task or room budget. Room lists also offer activity budget increases from cost chips.
Org LLM policy (BYOK)
A control-plane policy decides whether an organization may use the system model key or must bring its own; onboarding is blocked until this is resolved.
VM billing enforcement
A background job records VM compute, VM storage, and sandbox runtime charges. Sandbox temporary storage adds no separate charge; VM storage remains billable during hibernation. Organization-level enforcement can hibernate or delete VMs when charged usage exceeds the configured grace thresholds. Task, room, and agent admission checks do not interrupt running compute.