Together AI GPU pricing 2026
AI cloud that pairs H100 / H200 / B200 clusters with a Together Inference API for serving open-source models. Founded by the team behind RedPajama and the Open-LLaMA training runs. The cluster rate card publishes three columns per GPU: preemptible, on-demand, and reserved capacity in four commitment bands.
Published rate card
Per-GPU per-hour rates pulled from the vendor pricing page linked above, Last verified July 2026. Rates exclude storage, egress, and any managed-service uplift. Reserved-capacity contracts typically improve on these rates.
| GPU | Configuration | Per GPU per hour |
|---|---|---|
| H100 SXM | Reserved cluster, 91-180 day, per-GPU | $3.190 |
| H100 SXM | Reserved cluster, 31-90 day, per-GPU | $3.450 |
| H100 SXM | Reserved cluster, 7-30 day, per-GPU | $3.690 |
| H100 SXM | On-demand cluster, per-GPU | $3.990 |
| H100 SXM | Preemptible cluster, per-GPU | $1.990 |
| H200 SXM | Reserved cluster, 91-180 day, per-GPU | $3.990 |
| H200 SXM | On-demand cluster, per-GPU | $5.990 |
| H200 SXM | Preemptible cluster, per-GPU | $2.990 |
| B200 SXM | Reserved cluster, 91-180 day, per-GPU | $6.790 |
| B200 SXM | On-demand cluster, per-GPU | $8.190 |
| B200 SXM | Preemptible cluster, per-GPU | $4.090 |
Hidden costs to watch
- Reserved-cluster rates step down with term length across four bands (7-30, 31-90, 91-180 and 181+ days); the 181+ band is not published as a number, it is 'Contact us'.
- The preemptible column is the cheapest published rate but the capacity is interruptible, so it is not a substitute for a reserved cluster on a multi-day training run.
- Inference API billed per-million-tokens; serverless model hosting billed per-second.
- Network storage, dedicated endpoints, and fine-tuning runs all metered separately.
What Together AI is best for
Teams that want H100 / H200 / B200 clusters with managed networking and a high-throughput inference API alongside.
See the GPU cloud buying guideWorked example
Acme Vision Co. (illustrative example, not a real company) needs to train a 10 billion-parameter vision-language model on a fixed 8-GPU cluster for 30 days at 18 hours per day duty cycle.
At Together AI's cheapest published rate of $1.99 per GPU-hour, the run costs $8,597 for raw GPU compute, before storage, egress, and MLOps overhead. Add the typical 25 percent year-one uplift and the modelled spend is $10,746. Use the calculator on the homepage to model your own GPU class, cluster size, and duty cycle.
Last verified July 2026. Together AI rates change frequently. Always obtain a vendor quote before purchase.
Visit Together AI pricing page