Pricing

GPU clusters bill compute and storage separately. Compute pricing depends on the capacity type you choose (reserved, on-demand, or scheduled capacity), and preemptible nodes have their own discounted rate. Storage bills pay-as-you-go for as long as a volume exists.

Rates

All rates are in US dollars per GPU per hour. The duration columns show the reserved rate for each reservation length. The GPU clusters pricing table lists all GPU cluster hardware, including additional types available through sales.

Compute billing

• Reserved: Credits for the full reserved duration are charged upfront or deducted once the cluster is provisioned. Reserved rates are lower than on-demand rates for every reservation length. Usage beyond your reserved capacity is billed at on-demand rates.
• On-demand: You pay only for the time your cluster runs, with no upfront commitment.
• Scheduled capacity: Credits cover the full date window you select from your project’s capacity pool. The console shows an estimated total for the window before you create the cluster.
• Preemptible (Preview): Preemptible nodes, which you can add to Kubernetes and Slurm clusters alongside standard capacity, bill at a flat rate discounted from on-demand. Usage is metered every one to two minutes, so you pay only for the time a node is live. When you set a preemptible GPU target, the console shows the preemptible hourly rate and an estimated cost.

To create a cluster, you need enough credits to cover the full reservation, a minimum period of on-demand usage, four hours of the preemptible target, and storage for the same periods. Otherwise, the create request fails with a 403 error: Insufficient credits for cluster creation.

Storage billing

Storage is billed pay-as-you-go per GiB per month, based on each volume’s provisioned size. You can increase a volume’s size yourself, and the added capacity bills at the same rate. Decreasing a volume’s size requires contacting your account team, so start with a small volume and grow it as your data grows. Storage bills independently of the cluster lifecycle, so a volume keeps accruing charges even when no cluster is using it. Delete unused volumes to stop their charges.

Lifecycle policies

Clusters and storage volumes follow different lifecycle policies:

• Reserved clusters: A reserved cluster is decommissioned automatically when its reservation period ends. The cluster owner and project admins get email reminders seven days, three days, one day, and four hours before the reservation ends. To extend a reservation, open the cluster details page in the console and select Extend Reservation, or contact your account team.
• On-demand clusters: An on-demand cluster has no end date. It runs until you delete it or your credits run out.
• Scheduled capacity clusters: A scheduled capacity cluster is provisioned on the start date and decommissioned on the end date you select within the capacity pool window. If the pool window includes today, you can check Start now in the console to provision it immediately. If the pool has no end date, you can select No end instead, and the cluster runs until you delete it.
• Storage volumes: Volumes persist independently of clusters and remain available as long as your billing account is in good standing. Deleting a cluster doesn’t delete its volumes unless you choose to delete them with it. Your data persists as long as you use the static persistent volume Together provides.

Running out of credits

When your credits run out, each type of resource behaves differently:

• Reserved compute: Existing reservations stay active until their scheduled end date. On-demand capacity added to scale beyond the reservation is paused right away. If you don’t restore credits within the grace period (one hour by default), Together removes that capacity.
• On-demand compute: Together revokes access and pauses the cluster right away. If you don’t restore credits within the grace period, Together deletes the cluster.
• Preemptible compute: On reserved and scheduled capacity clusters, requests to raise the preemptible target fail with a 402 as soon as credits run out. When the grace period ends, Together sets the target to 0. Restoring credits lifts the 402, but the target stays at 0 until you raise it again. On an on-demand cluster, preemptible nodes are paused and deleted with the rest of the cluster.
• Storage volumes: Together revokes access right away. If you don’t restore credits within 72 hours, Together deletes the data.

When Together revokes access, it emails you the time left before deletion.

View usage and invoices

View your current usage, credits, and invoices in Settings > Billing. Each invoice has separate line items for reserved compute (upfront reservation charges), on-demand compute (hourly burst capacity usage), and storage (shared volume usage per GiB). Each line item shows the exact timeframe it covers.

FAQ

Can I get a refund for unused reservation time?

No. Reservations are non-refundable. The full reservation period is charged upfront and can’t be canceled or partially refunded.

Are there data ingress or egress fees?

No. Together AI doesn’t charge for data transfer into or out of your cluster, so you can move datasets, checkpoints, and results freely. You pay only for compute and storage.

Can I pause a cluster to save costs?

No. You can’t pause a reserved or on-demand cluster. A reserved cluster is charged upfront for the full reservation period. To stop compute charges on an on-demand cluster, delete it and recreate it later. Together pauses a cluster only when your credits run out.

When does my reservation start?

The reservation period starts when you create the cluster, not when it reaches Ready. The end date is the creation time plus the reservation length, so provisioning time counts toward the reservation.

Support

For billing questions, review your invoice in Settings > Billing or submit a support request. For reservation extensions, contact your account team.