Menu

Together AI GPU prices: H100, H200, B200 per hour

We may earn commissions when you shop through the links below. Learn more

Serverless open-model inference, dedicated endpoints and self-serve HGX GPU clusters.

GPU cloudInference hostingManaged AI hosting

GPU-focused cloud · Headquarters: San Francisco, United States

From published sources, not hands-on tested

From
$3.99≈ €3.56 (estimated conversion from $3.99)/GPU-hrH100 SXM 80GB · On-demand · rented as 8 GPUsChecked together.ai
GPU families
3
B200 · H200 · H100 · 3 models
Regions
5
4 in the EU
Billing
Per minute
Hourly, spot, reserved, serverless, usage-based

1 of 19 prices older than 7 days · Prices checked 42 sourcesSources (42)How we check

In short

Together AI (Together Computer, Inc., San Francisco) is an AI cloud that sells serverless inference billed per token or per output (20 chat, 29 image, 30 video and 9 audio models in its serverless catalog on 26 Sep 2026, mixing open-weight and proprietary partner models; no serverless embedding or rerank models), dedicated model endpoints billed per GPU-minute, managed fine-tuning, and NVIDIA HGX GPU clusters (H100, H200 and B200 self-serve; B300 also priced) run as Kubernetes or Slurm clusters in 8-GPU node increments.

Read full summary

On 26 Sep 2026 cluster GPUs listed at 3.99 to 9.99 USD per GPU-hour on demand, from 1.99 USD preemptible and from 3.19 USD with prepaid 91-180-day reservations, with shared storage at 0.16 USD per GiB-month and no cluster data-transfer fees.

From

$3.99≈ €3.56 (estimated conversion from $3.99)/GPU-hrH100 SXM 80GB · On-demand · rented as 8 GPUsChecked together.ai
Visit Together AI Opens together.ai in a new tab · Check the final plan, price and terms there before you order.Compare with alternatives to Together AI
Setup guides for Together AI
Billing increment
Per minute
Egress
No egress feesSource
Persistent storage
Yes
Managed service
Yes

Pricing

1 of 19 prices older than 7 days · Prices checked How we check

Billing: hourly, per minute, spot, reserved, serverless, usage-based. Lowest pay-as-you-go GPU price: $3.99≈ €3.56 (estimated conversion from $3.99) per GPU-hour (H100 SXM 80GB, on-demand, rented as 8 GPUs). Other price types start lower: $1.99≈ €1.78 (estimated conversion from $1.99) per GPU-hour (H100 SXM 80GB, spot, rented as 8 GPUs). Lowest plan price: no monthly fee (Serverless endpoint, usage-based); paid plans from $2,190.00≈ €1,956.58 (estimated conversion from $2,190.00) per month (Serverless endpoint).

Prices in USD, as listed, excluding VAT.≈ Estimated conversion from the listed USD price at the ECB reference rate of 23 Oct 2026 (1 EUR = 1.1193 USD). Billing currency not verified: providers may bill in USD, and your bank's rate and fees differ. Prices exclude VAT.

GPU offers17

Together AI GPU offers: 17 offers grouped by GPU; prices per GPU-hour, excl. VAT. Region: varies by offer. Availability: limited for every offer.

NVIDIA B200 SXM 180GB180 GB VRAM · 6 offers · from $4.09≈ €3.65 (estimated conversion from $4.09) per GPU-hour

Checked together.ai

  • Spot8-GPU node$4.09≈ €3.65 (estimated conversion from $4.09)/GPU-hrPer minute
  • Reserved8-GPU node$6.79≈ €6.07 (estimated conversion from $6.79)/GPU-hrHourly
  • Reserved8-GPU node$7.79≈ €6.96 (estimated conversion from $7.79)/GPU-hrHourly
  • Reserved8-GPU node$7.99≈ €7.14 (estimated conversion from $7.99)/GPU-hrHourly
  • On-demand8-GPU node$8.19≈ €7.32 (estimated conversion from $8.19)/GPU-hrHourly
  • On-demand$8.99≈ €8.03 (estimated conversion from $8.99)/GPU-hrPer minute
Show all 17 offers (from $1.99≈ €1.78 (estimated conversion from $1.99) per GPU-hour)

NVIDIA H200 SXM 141GB141 GB VRAM · 5 offers · from $2.99≈ €2.67 (estimated conversion from $2.99) per GPU-hour

Checked together.ai

  • Spot8-GPU node$2.99≈ €2.67 (estimated conversion from $2.99)/GPU-hrPer minute
  • Reserved8-GPU node$3.99≈ €3.56 (estimated conversion from $3.99)/GPU-hrHourly
  • Reserved8-GPU node$4.15≈ €3.71 (estimated conversion from $4.15)/GPU-hrHourly
  • Reserved8-GPU node$4.99≈ €4.46 (estimated conversion from $4.99)/GPU-hrHourly
  • On-demand8-GPU node$5.99≈ €5.35 (estimated conversion from $5.99)/GPU-hrHourly

NVIDIA H100 SXM 80GB80 GB VRAM · 6 offers · from $1.99≈ €1.78 (estimated conversion from $1.99) per GPU-hour

Checked together.ai

  • Spot8-GPU node$1.99≈ €1.78 (estimated conversion from $1.99)/GPU-hrPer minute
  • Reserved8-GPU node$3.19≈ €2.85 (estimated conversion from $3.19)/GPU-hrHourly
  • Reserved8-GPU node$3.45≈ €3.08 (estimated conversion from $3.45)/GPU-hrHourly
  • Reserved8-GPU node$3.69≈ €3.30 (estimated conversion from $3.69)/GPU-hrHourly
  • On-demand8-GPU node$3.99≈ €3.56 (estimated conversion from $3.99)/GPU-hrHourly
  • On-demand$5.49≈ €4.90 (estimated conversion from $5.49)/GPU-hrPer minute

Prices are per single GPU-hour, excluding VAT. Multi-GPU nodes are divided by their GPU count; the node size is shown with the price type. Each row shows when the price was observed and links to the page it was read from.

Research notes for these offers17
NVIDIA B200 SXM 180GB · Spot · 8-GPU node
NVIDIA HGX B200 (B200_SXM, 180 GB), preemptible compute: 4.09 USD per GPU-hour (32.72 USD per 8-GPU node-hour), 49.9% of the on-demand rate. Public preview for Kubernetes clusters only (not Slurm); flat discount, not a bid; metered and billed every one to two minutes (shown as per-minute billing). The preemptible target must be a multiple of 8 GPUs, is filled as spare capacity appears and is not guaranteed; nodes can be reclaimed at any time with a drain window of at most five minutes. Every cluster also needs at least one standard (on-demand or reserved) 8-GPU node. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA B200 SXM 180GB · Reserved · 8-GPU node
NVIDIA HGX B200 (B200_SXM, 180 GB), reserved 91-180 days: 6.79 USD per GPU-hour (54.32 USD per 8-GPU node-hour), 82.9% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA B200 SXM 180GB · Reserved · 8-GPU node
NVIDIA HGX B200 (B200_SXM, 180 GB), reserved 31-90 days: 7.79 USD per GPU-hour (62.32 USD per 8-GPU node-hour), 95.1% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA B200 SXM 180GB · Reserved · 8-GPU node
NVIDIA HGX B200 (B200_SXM, 180 GB), reserved 3–30 days (the pricing page's label for this tier, checked 20 Oct 2026; the GPU product pages still give a seven-day minimum): 7.99 USD per GPU-hour (63.92 USD per 8-GPU node-hour), 97.6% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA B200 SXM 180GB · On-demand · 8-GPU node
NVIDIA HGX B200 (B200_SXM, 180 GB), on-demand: 8.19 USD per GPU-hour (65.52 USD per 8-GPU node-hour). Billed per hour of cluster runtime with no commitment; how partial hours are rounded is not documented. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA B200 SXM 180GB · On-demand
NVIDIA HGX B200 180 GB (hardware ID 1xnvidia-b200-180gb): 8.99 USD per GPU-hour on-demand (pricing page and DMI docs agree). H200, B300 and GB300 dedicated endpoints are contact-sales only. Dedicated Model Inference (single-tenant GPU replicas serving a Together-hosted, fine-tuned or uploaded model through the same inference API as serverless; no host access is documented). Billed by the minute per ready replica; provisioning, cold starts and replicas that are not ready are not billed, and a deployment scaled to zero or stopped costs nothing. Multi-GPU configs cost proportionally more (a 4-GPU config costs 4x). Placement regions are controlled by placement profiles; live regions and capacity headroom are only returned by an authenticated API. Reserved (monthly) capacity is quoted by sales.
NVIDIA H200 SXM 141GB · Spot · 8-GPU node
NVIDIA HGX H200 (H200_SXM; Together's GPU cards say 140 GB, its DMI docs say 141 GB), preemptible compute: 2.99 USD per GPU-hour (23.92 USD per 8-GPU node-hour), 49.9% of the on-demand rate. Public preview for Kubernetes clusters only (not Slurm); flat discount, not a bid; metered and billed every one to two minutes (shown as per-minute billing). The preemptible target must be a multiple of 8 GPUs, is filled as spare capacity appears and is not guaranteed; nodes can be reclaimed at any time with a drain window of at most five minutes. Every cluster also needs at least one standard (on-demand or reserved) 8-GPU node. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H200 SXM 141GB · Reserved · 8-GPU node
NVIDIA HGX H200 (H200_SXM; Together's GPU cards say 140 GB, its DMI docs say 141 GB), reserved 91-180 days: 3.99 USD per GPU-hour (31.92 USD per 8-GPU node-hour), 66.6% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H200 SXM 141GB · Reserved · 8-GPU node
NVIDIA HGX H200 (H200_SXM; Together's GPU cards say 140 GB, its DMI docs say 141 GB), reserved 31-90 days: 4.15 USD per GPU-hour (33.20 USD per 8-GPU node-hour), 69.3% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H200 SXM 141GB · Reserved · 8-GPU node
NVIDIA HGX H200 (H200_SXM; Together's GPU cards say 140 GB, its DMI docs say 141 GB), reserved 3–30 days (the pricing page's label for this tier, checked 20 Oct 2026; the GPU product pages still give a seven-day minimum): 4.99 USD per GPU-hour (39.92 USD per 8-GPU node-hour), 83.3% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H200 SXM 141GB · On-demand · 8-GPU node
NVIDIA HGX H200 (H200_SXM; Together's GPU cards say 140 GB, its DMI docs say 141 GB), on-demand: 5.99 USD per GPU-hour (47.92 USD per 8-GPU node-hour). Billed per hour of cluster runtime with no commitment; how partial hours are rounded is not documented. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H100 SXM 80GB · Spot · 8-GPU node
NVIDIA HGX H100 (H100_SXM, 80 GB HBM3), preemptible compute: 1.99 USD per GPU-hour (15.92 USD per 8-GPU node-hour), 49.9% of the on-demand rate. Public preview for Kubernetes clusters only (not Slurm); flat discount, not a bid; metered and billed every one to two minutes (shown as per-minute billing). The preemptible target must be a multiple of 8 GPUs, is filled as spare capacity appears and is not guaranteed; nodes can be reclaimed at any time with a drain window of at most five minutes. Every cluster also needs at least one standard (on-demand or reserved) 8-GPU node. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H100 SXM 80GB · Reserved · 8-GPU node
NVIDIA HGX H100 (H100_SXM, 80 GB HBM3), reserved 91-180 days: 3.19 USD per GPU-hour (25.52 USD per 8-GPU node-hour), 79.9% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H100 SXM 80GB · Reserved · 8-GPU node
NVIDIA HGX H100 (H100_SXM, 80 GB HBM3), reserved 31-90 days: 3.45 USD per GPU-hour (27.60 USD per 8-GPU node-hour), 86.5% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H100 SXM 80GB · Reserved · 8-GPU node
NVIDIA HGX H100 (H100_SXM, 80 GB HBM3), reserved 3–30 days (the pricing page's label for this tier, checked 20 Oct 2026; the GPU product pages still give a seven-day minimum): 3.69 USD per GPU-hour (29.52 USD per 8-GPU node-hour), 92.5% of the on-demand rate. Charged upfront for the full reservation (GPUs x hours x rate) once the cluster is provisioned; non-refundable and cannot be paused; clusters are decommissioned when the reservation ends unless extended; usage beyond the reservation is billed at on-demand rates. The docs describe self-serve reservations of 1-90 days, while the pricing page publishes tiers to 180 days (181+ days: contact); the pricing page's shortest tier starts at 3 days (20 Oct 2026); no rate is published for shorter reservations. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H100 SXM 80GB · On-demand · 8-GPU node
NVIDIA HGX H100 (H100_SXM, 80 GB HBM3), on-demand: 3.99 USD per GPU-hour (31.92 USD per 8-GPU node-hour). Billed per hour of cluster runtime with no commitment; how partial hours are rounded is not documented. Together GPU Clusters (HGX, SXM, InfiniBand; Kubernetes or Slurm-on-Kubernetes). Sold only in multiples of 8 GPUs (the GPU count must be a multiple of 8), so the smallest purchase is one 8-GPU node; price shown per GPU-hour as published. Node vCPU, RAM and local NVMe sizes are not published. Region is chosen from live inventory (default 'Any region'); the region list is only visible after sign-in (console region picker or the authenticated regions API). Capacity is subject to real-time availability (docs: 'Notify Me' when unavailable; capacity requests not guaranteed). No ingress or egress fees; shared storage billed separately.
NVIDIA H100 SXM 80GB · On-demand
NVIDIA HGX H100 80 GB (hardware ID 1xnvidia-h100-80gb): list price 5.49 USD per GPU-hour, as the pricing page shows it on 1 Oct 2026. A promotion lowered it to 3.99 USD per GPU-hour until 30 Sep 2026 (on 26 Sep 2026 the DMI docs hardware table listed 3.99 USD); the docs' own cost example uses 5.49 USD (about 132 USD per day). Dedicated Model Inference (single-tenant GPU replicas serving a Together-hosted, fine-tuned or uploaded model through the same inference API as serverless; no host access is documented). Billed by the minute per ready replica; provisioning, cold starts and replicas that are not ready are not billed, and a deployment scaled to zero or stopped costs nothing. Multi-GPU configs cost proportionally more (a 4-GPU config costs 4x). Placement regions are controlled by placement profiles; live regions and capacity headroom are only returned by an authenticated API. Reserved (monthly) capacity is quoted by sales.

Compare these prices with other providers’: H100 prices comparedH200 prices compared.

Plans3

Serverless endpoint2 plans · no monthly fee (usage-based) · paid from $2,190.00≈ €1,956.58 (estimated conversion from $2,190.00) per month

  • Serverless Inference (pay as you go)Global · ManagedChecked · Staletogether.aiNo monthly fee
  • Provisioned Throughput (per PTU)ManagedChecked docs.together.ai$2,190.00≈ €1,956.58 (estimated conversion from $2,190.00)/monthor $3.00≈ €2.68 (estimated conversion from $3.00)/hr

CPU instance1 plan · price not verified

  • Code Sandbox (VM sandboxes)ManagedChecked · Staletogether.aiNot verified
Research notes for these plans3
Serverless Inference (pay as you go)
No subscription or platform fee was found for self-serve use; usage is billed per token (per image, video, audio minute or character for other modalities) from prepaid credits: minimum credit purchase 15 USD for an organization's first purchase and 5 USD after that (docs, 14 Oct 2026), no free trial, credits do not expire, API access stops at a zero balance (credits docs). Chat examples on 26 Sep 2026, USD per 1M input / output tokens: gpt-oss-120B 0.15 / 0.60; Qwen3.8 Flash 0.09 / 0.28; Llama 3.3 70B 1.04 / 1.04; DeepSeek V4 Pro 0813 1.32 / 3.96 (0.13 cached input); Kimi K3 3.00 / 15.00 (0.30 cached input). The docs index describes a Batch API at up to 50% lower cost. Best-effort service with dynamic per-model rate limits and no region selection. Scale and Enterprise contracts are billed under their own terms. The serverless catalog in the docs lists 20 chat, 3 vision (a subset of the chat models), 29 image, 30 video and 9 audio models on 26 Sep 2026 and states there are currently no embedding, rerank or moderation models on serverless; Together's docs index still describes '100+ open-source models'. The pricing page's 'Batch API price' toggle showed no batch prices on 26 Sep 2026.
Provisioned Throughput (per PTU)
Sales-only reserved inference capacity (docs: 'currently unavailable for self-service') with a fixed term of one month or more, priced at a flat 0.05 USD per provisioned throughput unit (PTU) per minute (docs and pricing page). Hourly figure = 0.05 x 60; monthly figure = 0.05 x ~43,800 minutes, the 24/7 assumption stated on the pricing page (its calculator example of 353 PTUs = 773,070 USD per month matches 2,190 USD per PTU-month). Throughput per PTU is model-specific, in tokens per minute (input / cached / output): MiniMax M3 166,667 / 833,333 / 41,667; Kimi K3 16,667 / 166,667 / 3,333; GLM-5.2 35,714 / 192,308 / 11,364. SLA: throughput up to the purchased PTUs and at least 99% of eligible requests succeeding each month. Unused PTUs are not refunded; overage may fall back to serverless at standard rates without SLA. Real commitments buy many PTUs.
Code Sandbox (VM sandboxes)
Metered per resource rather than per instance: 0.0446 USD per vCPU-hour and 0.0149 USD per GiB RAM-hour (pricing page), so no fixed monthly or hourly price applies. Firecracker microVMs (CodeSandbox infrastructure) with memory snapshot/restore, filesystem persistence and Docker / Dev Container support. Together's docs say Code Sandbox is on Together custom plans (contact sales), with self-serve access through CodeSandbox SDK plans (VM credits at 0.01486 USD; for example Nano, 2 cores and 4 GB, 0.1486 USD per hour; billed per started minute). The Code Interpreter API is priced at 0.03 USD per 60-minute session and is self-serve with a standard Together API key (code interpreter docs).

Storage and data transfer

GPU Cluster shared volumes (Shared Filesystem)
$0.149≈ €0.133 (estimated conversion from $0.149) per GB-month
File storage · Persistent0.16 USD per GiB-month (quoted per GiB; about 0.149 USD per decimal GB). Pricing page: Together AI Pricing (opens in a new tab)
MoreRemote-attached, high-throughput parallel filesystem (Together names WEKA and VAST) shared by all nodes; minimum volume 1 TiB (about 163.84 USD per month at list price), sized in whole TiB, self-serve increases only (decreases via the account team). Volumes are region-specific, persist after cluster deletion, keep accruing charges until deleted, and move to on-demand pricing when a reservation ends. Mounted as a static PersistentVolume on Kubernetes and at /home on Slurm. Published as USD 0.16 per GiB-month; stored per GB (1 GiB = 1.073741824 GB).
Node-local NVMe (GPU Clusters)
Not verified
Local NVMe · Not persistentIncluded with each GPU node; capacity per node is not published. Docs: Cluster storage | Together AI docs (opens in a new tab)
MoreEphemeral: data can be lost during node migrations, recreations, maintenance or rescheduling. /home on Kubernetes clusters is also node-local and ephemeral.
Managed Storage (object storage and parallel filesystems)
Not verified
Object storage · PersistentProvisioned and operated by Together on request (WEKA or VAST), with zero egress fees, encryption in transit and at rest, HIPAA-aligned options and regions in North America, Europe or Asia/Middle East. Docs: Managed Storage | Together AI (opens in a new tab)
MoreNo public per-GB price. The pricing page's 'Managed Storage' navigation link leads to its Storage table, which lists only the Shared Filesystem (0.16 USD per GiB-month); no price is published for object storage or for dedicated WEKA or VAST storage clusters.
Egress (data out)
$0.00≈ €0.00 (estimated conversion from $0.00) per GB
GPU Clusters billing FAQ: 'Together AI does not charge for data transfer into or out of your cluster'. Docs: Billing & pricing (GPU clusters) | Together AI docs (opens in a new tab)
MoreManaged Storage also advertises zero egress fees. Serverless, dedicated endpoints and provisioned throughput are billed per token, minute or PTU with no separate data-transfer charges documented.

Before you compare

From Together AI’s record: its first stated uses and limitation, how its prices are set, and the providers compared below.

Suited to
  • Applications that want OpenAI-compatible, per-token access to many open-weight models without managing GPUs
  • Multi-node training or fine-tuning on 8-GPU HGX H100, H200 or B200 nodes with Kubernetes or Slurm over InfiniBand
All 4 uses it suits
Main limitation
GPU clusters are sold in multiples of 8 GPUs, so even small experiments pay for at least one full 8-GPU node
All 11 limitations
Price basis
GPU prices per GPU-hour, billed per minute (lowest price rented as 8 GPUs); egress included; 3 storage options listed
All prices, storage and egress
Alternatives
Modal provider pageHyperstack provider pageLambda provider pageHetzner provider page
Compare side by side with Together AIHow they differ

Who it fits

Based on Together AI’s public documentation, checked . We don’t score providers. How we build records

Suited to

  • Applications that want OpenAI-compatible, per-token access to many open-weight models without managing GPUs
  • Multi-node training or fine-tuning on 8-GPU HGX H100, H200 or B200 nodes with Kubernetes or Slurm over InfiniBand
  • Jobs with a known duration that can use published, prepaid reserved GPU rates for 7 to 180 days
  • Serving catalog or fine-tuned open models on dedicated GPUs billed per minute with autoscaling and scale-to-zero

Good fit for

  • TrainingStrong fitSelf-serve 8-GPU HGX H100, H200 and B200 nodes with InfiniBand, Kubernetes or Slurm, shared parallel-filesystem volumes and published reserved rates; Together cites up to 256 GPUs on demand and 4,000+ GPUs reserved. Capacity depends on availability, and reservations are prepaid and non-refundable.
  • Fine-tuningStrong fitManaged fine-tuning (LoRA and full, SFT and DPO) priced per training token for listed open models with per-job minimum charges, plus GPU clusters and preemptible nodes for custom checkpointed runs; fine-tuned models deploy to dedicated endpoints.
  • Batch inferenceStrong fitA Batch API for asynchronous jobs (Together cites up to 50% lower cost than serverless) and preemptible cluster nodes, which Together recommends for checkpointed batch inference at about half the on-demand rate.
  • Real-time inferenceStrong fitOpenAI-compatible serverless API (20 chat models in the serverless catalog on 26 Sep 2026) plus dedicated endpoints with autoscaling replicas billed per minute; provisioned throughput adds a request-success SLA through sales. Serverless is best-effort with dynamic rate limits and no region selection.
Show 2 more
  • AI agentsSuitableTogether's docs cover OpenClaw, LangGraph, CrewAI and other agent frameworks with Together as the model provider, Hermes Agent's docs include a Together config, and a code interpreter API runs model-generated code (0.03 USD per 60-minute session). Agent gateways themselves must be hosted elsewhere; the code interpreter API is self-serve, while Together-billed VM sandboxes are on custom (sales) plans or self-serve only through a separate CodeSandbox account.
  • Private LLM hostingSuitableDedicated endpoints run on single-tenant GPUs, GPU clusters give full cluster control, zero data retention can be enabled and training on customer data is opt-in. Prompts and responses are stored by default until ZDR is enabled, passthrough models forward data to third parties unless disabled, and EU or VPC deployments go through sales. The standard Terms of Service (updated 19 May 2026, section 4.1) also bar sending financial, medical or other sensitive personal data through the Services; how this interacts with the HIPAA-attested placement in the dedicated-inference API and HIPAA-aligned Managed Storage is not documented.

Look elsewhere for

  • RAG applicationsPartial fitChat models are served through the same OpenAI-compatible API, but on 26 Sep 2026 the serverless catalog lists no embedding or rerank models; Together says rerank models such as mxbai-rerank-large-v2 are only available on dedicated inference, whose published catalog shows only chat models. Embeddings, the vector database and application hosting must come from elsewhere or be self-hosted on a GPU cluster (minimum one 8-GPU node).
  • RenderingNot suitableOnly HGX H100, H200, B200 and B300 have published self-serve prices, sold as 8-GPU nodes; no rendering-oriented instances, templates or remote-desktop sessions are documented (L40 and RTX 6000 GPU types appear in the cluster API without public prices).

Not sure Together AI fits what you want to build?Help me choose

Strengths and limitations

Summarised from the research record for Together AI. We don’t score providers.

Strengths

  • Published per-GPU-hour prices for on-demand, preemptible and reserved (3–30, 31–90, 91–180 days) cluster capacity
  • Preemptible cluster nodes at about half the on-demand rate (H100 1.99 vs 3.99 USD per GPU-hour), metered every one to two minutes
  • No ingress or egress fees for GPU clusters; shared volumes at 0.16 USD per GiB-month persist independently of clusters
  • Managed Kubernetes or Slurm-on-Kubernetes clusters with SSH access, InfiniBand, autoscaling, health checks and node repair
Show 3 more
  • One OpenAI-compatible inference API across serverless, batch, provisioned throughput and dedicated endpoints
  • Dedicated endpoints bill per minute only for ready replicas and cost nothing when scaled to zero
  • SOC 2 Type 2 report and ISO 27001:2022 certificate via the trust center; zero data retention option; training on customer data is opt-in

Limitations

  • GPU clusters are sold in multiples of 8 GPUs, so even small experiments pay for at least one full 8-GPU node
  • Published cluster prices cover only HGX H100, H200, B200 and B300; GB200/GB300 NVL72 clusters and dedicated-endpoint H200, B300, GB200 and GB300 are contact-sales
  • Reservations are prepaid and non-refundable; the docs describe self-serve reservations of 1-90 days while published tiers run to 180 days
  • Preemptible compute is a public preview for Kubernetes clusters only, with no capacity guarantee, a five-minute maximum drain window and a required standard node per cluster
Show 7 more
  • Serverless inference has no region selection, EU deployments are arranged on request, and the cluster region list is only visible after sign-in (console region picker or the authenticated regions API)
  • The H100 dedicated-endpoint rate of 3.99 USD per GPU-hour is a promotion ending 30 Sep 2026; the list price is 5.49 USD
  • Fully prepaid platform with a 5 USD minimum credit purchase and no free trial; API access stops at a zero balance
  • No public uptime SLA: the Terms of Service make no uptime guarantee unless one is agreed in an order form, and the only documented SLA (99% of eligible requests) covers sales-only provisioned throughput
  • Custom Docker containers on managed inference (Dedicated Container Inference) are sales-only, and Code Sandbox VMs are on Together custom plans (sales) or self-serve only through a separate CodeSandbox account
  • No serverless embedding, rerank or moderation models on 26 Sep 2026; Together says rerank models run only on dedicated inference, whose published catalog lists only chat models
  • The standard Terms of Service (updated 19 May 2026, section 4.1) bar sending financial, medical or other sensitive personal data through the Services

Infrastructure

Regions

5 regions across 5 countries; 4 are recorded as EU regions.

Europe (4)
France (EU); Netherlands (EU); Romania (EU); Sweden (EU)
North America (1)
United States (multiple data centers)

Sources: Pricing page: NVIDIA GPU Clusters: H100, H200, B200, GB200 | Together AI (opens in a new tab), Secondary source: Together AI Continues European Expansion: Infrastructure Now Live and Operational in Sweden (opens in a new tab)

Region notes5
United States (multiple data centers)
Together cites '2GW+ in the portfolio with 600MW of near-term capacity in US'; its privacy doc says models such as DeepSeek are hosted in North America data centers. The GPU Clusters page also says 'Launch close to your users and data across 25+ cities' without naming them. Region identifiers and data-center cities are not published; the live list is only visible after sign-in (console region picker or the authenticated regions API; docs examples such as us-central-8 are illustrative, per Together). AIHostCompare lists this capacity as one United States region.
Sweden
Together press release (4 Sep 2025): GPU infrastructure in Sweden is live and powers serverless inference; GPU Clusters and Dedicated Endpoints can be requested in Sweden via sales forms. Also listed in the European capacity block on the GPU Clusters page (Sep 2026). Self-serve availability in the cluster console not verified; city and region code not published.
France
Named in '150 MW+ available in Europe: France, Netherlands, Sweden, and Romania' (GPU Clusters and H100 pages), in the H100 page FAQ and as a priority deployment in the June 2025 Europe blog. Access appears to be by request or sales; city, region code and self-serve availability not published.
Netherlands
Named in the European capacity block on the GPU Clusters and H100 pages only. City, region code, products and self-serve availability not published; treat as sales-led capacity.
Romania
Named in the European capacity block on the GPU Clusters and H100 pages only. City, region code, products and self-serve availability not published; treat as sales-led capacity. The H100 page FAQ instead lists UK, Spain, France, Portugal and Iceland for Europe, and the June 2025 blog named France, the UK, Italy and Portugal as priority sites.

Frameworks

How Together AI works with common AI serving, agent and training frameworks. Support levels, strongest first: an official template, documentation, a community guide, or manual setup on a general-purpose machine.

Hosting

Support for running the framework itself on Together AI.

PyTorch
Documented setup
Cluster docs run GPU pods from pytorch/pytorch CUDA images; the SkyPilot example installs torch; the preemptible docs cite Ray Train and PyTorch Lightning checkpoint restarts; the cluster API offers a 'Model Aware TorchPass' add-on. Docs: Cluster management | Together AI docs (opens in a new tab)
Ollama
Manual setup
No Together guide for Ollama. It could run as a container on a GPU Cluster, but the smallest unit is an 8-GPU HGX node, far larger than typical Ollama deployments; untested. Docs: Cluster management | Together AI docs (opens in a new tab)
SGLang
Manual setup
No Together guide for SGLang; self-managed on GPU Clusters as a container workload (minimum one 8-GPU node), untested. Docs: Cluster management | Together AI docs (opens in a new tab)
Text Generation Inference
Manual setup
No Together guide for TGI; self-managed on GPU Clusters as a container workload (minimum one 8-GPU node), untested. Docs: Cluster management | Together AI docs (opens in a new tab)
vLLM
Manual setup
No Together guide for vLLM. Docs: Cluster management | Together AI docs (opens in a new tab)
MoreGPU Clusters run arbitrary CUDA container images as Kubernetes pods, so a self-managed vLLM server is possible but untested (minimum one 8-GPU node). Dedicated endpoints use Together's own inference stack, not a user-selectable vLLM; Dedicated Container Inference (sales-only) runs custom images through the Sprocket SDK.

Model API

Together AI serves the models the framework calls; the framework itself runs elsewhere.

Dify
Documented setup
Official Dify Marketplace model plugin langgenius/togetherai (v0.0.9, 'Verified by Dify') adds Together models to Dify. Dify itself runs on Dify Cloud or your own server; Together does not host it. Docs: Together.AI - Dify Marketplace (opens in a new tab)
Flowise
Documented setup
Flowise docs list a ChatTogetherAI chat-model node (its page is a 'work in progress' stub). Flowise runs on your own server; Together supplies the models. Docs: ChatTogetherAI | FlowiseAI (opens in a new tab)
Hermes Agent
Documented setup
Hermes Agent's provider docs list Together AI as an OpenAI-compatible provider and include a Together AI recipe in its 'Cookbook: Together AI, Groq, Perplexity' section (providers.together with key_env TOGETHER_API_KEY). Docs: LLM and Model Providers | Hermes Agent (opens in a new tab)
MoreHermes runs on your own host; Together supplies the models.
LangGraph
Documented setup
Together docs page with LangGraph examples (langchain-together in Python, ChatTogetherAI in TypeScript) and two cookbook notebooks. Docs: LangGraph | Together AI docs (opens in a new tab)
MoreLangGraph runs in your own environment; Together supplies the models.
OpenClaw
Documented setup
Together documents OpenClaw with Together as the model provider ('openclaw onboard --auth-choice together-api-key', models prefixed together/). Docs: Configure OpenClaw with Together AI models | Together AI docs (opens in a new tab)
MoreThe OpenClaw gateway runs on your own machine or server; Together does not host it.
n8n
Community guide
No official Together node or credential in n8n and no Together guide for n8n. Secondary source: Generate AI videos from scripts with DeepSeek, TTS, and Together.ai | n8n workflow template (opens in a new tab)
MoreCommunity evidence: n8n's workflow gallery lists the community-created template 'Generate AI videos from scripts with DeepSeek, TTS, and Together.ai' (creator Lakindu Siriwardana), which calls Together's API for image generation through HTTP Request nodes with a Together API key, and in the n8n forum thread (Feb to Jun 2025) a community Top Supporter explains calling together.ai with a Generic Bearer Auth credential. n8n runs on your own host or n8n Cloud; Together supplies the models.

Together AI is not listed on the Ollama hosting, vLLM hosting, Hermes Agent hosting and OpenClaw hosting pages. These pages list providers that host the framework with an official template, documentation or a community guide.

Operations

Managed service: yes. Each feature below states what the provider documents, with the source it was checked against. “Not verified” means no reliable source was found.

Features

Compute and access

Custom containers (Docker)
Yes
GPU Clusters run arbitrary container images as Kubernetes pods (docs use pytorch/pytorch CUDA images). Docs: Cluster management | Together AI docs (opens in a new tab)
MoreCustom Docker images on managed inference (Dedicated Container Inference) and Docker-based Code Sandbox environments are sales-gated.
Root access
Yes
GPU Clusters only: direct SSH to Slurm login/compute nodes and Kubernetes worker nodes, and the docs have users run 'sudo scontrol' on Slurm nodes; an admin kubeconfig is provided. Docs: Cluster management | Together AI docs (opens in a new tab)
MoreServerless, dedicated endpoints and provisioned throughput have no host access. On Kubernetes clusters, worker-node SSH is for monitoring and debugging (no direct GPU access; GPU work runs in pods); sudo appears only in the Slurm instructions.
SSH access
Yes
SSH to Slurm login and compute nodes (OIDC via the Together CLI or key-based through a jump host) and to Kubernetes worker nodes for monitoring and debugging. Docs: Cluster management | Together AI docs (opens in a new tab)
API access
Yes
REST API, Python and TypeScript SDKs and the tg CLI for clusters, endpoints and deployments (Terraform support mentioned in the API reference); inference API is OpenAI-compatible. Docs: API & integrations (GPU clusters) | Together AI docs (opens in a new tab)
Jupyter notebooks
No
No hosted notebook product appears in the documentation index; Jupyter can be self-installed on a GPU cluster (not documented). Docs: Together AI docs index (opens in a new tab)
Templates
Partial
Dedicated endpoints deploy catalog models with one CLI command, and Code Sandbox offers a template library. No one-click application templates (for example n8n or Dify) were found. Docs: Dedicated model inference overview | Together AI docs (opens in a new tab)
Multi-GPU instances
Yes
Clusters start at one 8-GPU HGX node and scale to 256 GPUs on demand or 4,000+ GPUs reserved over InfiniBand; dedicated endpoints support multi-GPU configs. Pricing page: NVIDIA GPU Clusters: H100, H200, B200, GB200 | Together AI (opens in a new tab)
Bare metal
Partial
Together describes each cluster as 'dedicated bare metal' and advertises 'bare-metal performance', but the cluster API's remediation modes refer to GPU worker VMs on bare-metal hosts, and nodes are delivered as managed Kubernetes or Slurm workers. Pricing page: NVIDIA HGX H100 Cluster Pricing & Specs | Together AI (opens in a new tab)
MoreNo standalone bare-metal server product is documented.
Kubernetes
Yes
Each GPU cluster is a Kubernetes cluster (kubeadm-based per the product page) with kubectl and dashboard access; Slurm clusters also expose a kubeconfig. Docs: GPU clusters overview | Together AI docs (opens in a new tab)
MoreOIDC sign-in and Kubernetes RBAC are documented.
Show all 21 features

Management

Managed service
Yes
Fully managed serverless inference, dedicated endpoints and fine-tuning; GPU Clusters come with managed Kubernetes (HA control plane, managed upgrades) or Slurm-on-Kubernetes, health checks and node repair. Pricing page: NVIDIA GPU Clusters: H100, H200, B200, GB200 | Together AI (opens in a new tab)
Serverless
Yes
Serverless inference billed per token or per output unit (20 chat, 29 image, 30 video and 9 audio models in the serverless catalog on 26 Sep 2026; the docs index still says '100+ open-source models'); dedicated endpoints can also scale to zero replicas and then cost nothing. Docs: Available models (serverless) | Together AI docs (opens in a new tab)

Storage and data

Persistent storage
Yes
Shared volumes persist independently of clusters (0.16 USD per GiB-month, 1 TiB minimum); /home is persistent on Slurm clusters. Docs: Cluster storage | Together AI docs (opens in a new tab)
Snapshots
Partial
Code Sandbox microVMs support memory snapshot/restore and resume from snapshot (sales-gated at Together). No snapshot feature is documented for GPU Cluster shared volumes or nodes. Docs: Code sandbox | Together AI docs (opens in a new tab)
Managed backups
Not verified
No user-facing backup or restore service for shared volumes is documented; the trust center lists internal continuity and disaster-recovery controls only.

Billing

Per-second billing
No
Dedicated endpoints bill per minute, on-demand and reserved clusters per hour, preemptible nodes every one to two minutes; serverless is per token. Docs: Dedicated model inference pricing | Together AI docs (opens in a new tab)
Spot pricing
Yes
Preemptible cluster nodes (public preview, Kubernetes only) at a flat discount, about 50% of on-demand (H100 1.99 vs 3.99 USD per GPU-hour). Docs: Preemptible compute | Together AI docs (opens in a new tab)
Reserved pricing
Yes
Published reserved GPU-cluster rates for 3–30, 31–90 and 91–180 days (181+ days on request); prepaid and non-refundable. Pricing page: Together AI Pricing (opens in a new tab)
MoreReserved dedicated-endpoint capacity and provisioned throughput are quoted by sales.

Regions and compliance

EU region
Partial
Together states 150 MW+ of European capacity in France, Netherlands, Sweden and Romania, and its Sweden infrastructure has been live since Sep 2025 for serverless inference, with GPU Clusters and Dedicated Endpoints available on request. Docs: Privacy and security | Together AI docs (opens in a new tab)
MoreServerless endpoints offer no region selection, EU-region and VPC deployments are arranged for enterprise customers, and whether an EU region appears in the self-serve cluster console could not be verified (the region list is only visible after sign-in). Together's own pages list different European countries (see regions notes).
SOC 2 report
Yes
SOC 2 Type 2 report available on request via the trust center (updated 17 Jun 2026); ISO 27001:2022 certified (announced 29 May 2026). Docs: Together.ai Trust Center (opens in a new tab)
GDPR data processing agreement
Yes
Data Processing Addendum available on request via the trust center (contents not publicly readable). Docs: Together.ai Trust Center (opens in a new tab)
MoreThe privacy policy's EEA/UK section cites EU standard contractual clauses and data-subject rights; zero data retention can be enabled.
Customer-managed keys (BYOK)
Not verified
No customer-managed encryption keys are documented. Dedicated Container Inference has a secrets API for application credentials.

SLA and support

Uptime SLA
None published
Terms of Service (updated 19 May 2026), section 1.2: 'No guarantees are made with respect to the Services' quality, stability, uptime, or reliability, unless otherwise agreed between the parties in an order form.' No uptime SLA is published for serverless (described as best-effort), dedicated endpoints or GPU clusters. Docs: Terms of Service | Together AI (opens in a new tab)
MoreProvisioned throughput (sales-only, one-month minimum) carries an SLA of at least 99% of eligible requests completing successfully each month plus a throughput commitment, which is a request-success target rather than uptime. The status page shows 90-day uptime per serverless model (98.729% to 100% at 18:49 UTC on 26 Sep 2026, and 98.728% at 18:04 UTC, because the window rolls), which is historical, not a commitment.
Support channels
4 listed
  • Support portal with ticket submission (support.together.ai)
  • Email (support@together.ai)
  • Discord community
  • Shared Slack channel (organizations that have one)
No published response-time targets were found. Security reports go through a private HackerOne program (privacy and security docs). Docs: Support | Together AI docs (opens in a new tab)

Sources and verification

42 sources, accessed between 26 Sep 2026 and 20 Oct 2026. Official provider pages are listed first; the 4 secondary sources are labelled as such.

How we verify

Prices, regions, storage, features, framework support, SLA and support channels link to the page they were read from, and every price shows when it was checked. Fit labels, strengths and limitations summarise the same research record. The Together AI record was last checked on 26 Sep 2026. A check date older than 7 days is marked “Stale” with an amber dot, and one older than 30 days “Needs refresh”.

Last verified
26 Sep 2026, 19:04 UTC
Checked · Stale
Method
Source-by-source review by AI agents, independently re-checked
Confidence
High
Checked by
AI research agent; every value re-checked against its source by a separate AI agent

Verification notes

On 26 September 2026 we checked every price in this listing against Together AI's live pricing page, documentation, model catalogs and terms. Where the H100 page's FAQ contradicts its own price and region tables, the listing follows the tables. Not verified: cluster node sizes, the self-serve region list (it needs sign-in) and partial-hour rounding. Later corrections are listed on the methodology page.

Commercial relationship

AIHostCompare has no active affiliate link with Together AI; the Visit provider link goes to its official website. Commissions never change ordering, alternatives, fit labels or conclusions. Read the affiliate disclosure.

Pricing

  1. 1.NVIDIA GPU Clusters: H100, H200, B200, GB200 | Together AI (opens in a new tab) · Official pricing page · accessed
  2. 2.NVIDIA HGX H100 Cluster Pricing & Specs | Together AI (opens in a new tab) · Official pricing page · accessed
  3. 3.Together AI Pricing (opens in a new tab) · Official pricing page · accessed

Documentation

  1. 4.API & integrations (GPU clusters) | Together AI docs (opens in a new tab) · Official docs · accessed
  2. 5.Available models (serverless) | Together AI docs (opens in a new tab) · Official docs · accessed
  3. 6.Billing & pricing (GPU clusters) | Together AI docs (opens in a new tab) · Official docs · accessed
  4. 7.Bringing 100,000 GPUs to Europe | Together AI (opens in a new tab) · Official docs · accessed
  5. 8.ChatTogetherAI | FlowiseAI (opens in a new tab) · Official docs · accessed
  6. 9.Cluster management | Together AI docs (opens in a new tab) · Official docs · accessed
  7. 10.Cluster storage | Together AI docs (opens in a new tab) · Official docs · accessed
Show all 42 sources

Documentation

  1. 11.Code interpreter | Together AI docs (opens in a new tab) · Official docs · accessed
  2. 12.Code sandbox | Together AI docs (opens in a new tab) · Official docs · accessed
  3. 13.Configure OpenClaw with Together AI models | Together AI docs (opens in a new tab) · Official docs · accessed
  4. 14.Create a GPU cluster (API reference) | Together AI docs (opens in a new tab) · Official docs · accessed
  5. 15.Credits | Together AI docs (opens in a new tab) · Official docs · accessed
  6. 16.Dedicated Container Inference overview | Together AI docs (opens in a new tab) · Official docs · accessed
  7. 17.Dedicated model inference overview | Together AI docs (opens in a new tab) · Official docs · accessed
  8. 18.Dedicated model inference pricing | Together AI docs (opens in a new tab) · Official docs · accessed
  9. 19.GPU clusters overview | Together AI docs (opens in a new tab) · Official docs · accessed
  10. 20.GPU clusters quickstart | Together AI docs (opens in a new tab) · Official docs · accessed
  11. 21.LangGraph | Together AI docs (opens in a new tab) · Official docs · accessed
  12. 22.List inference instance types (API reference) | Together AI docs (opens in a new tab) · Official docs · accessed
  13. 23.List placement profiles (API reference) | Together AI docs (opens in a new tab) · Official docs · accessed
  14. 24.LLM and Model Providers | Hermes Agent (opens in a new tab) · Official docs · accessed
  15. 25.Managed Storage | Together AI (opens in a new tab) · Official docs · accessed
  16. 26.OpenAI credentials | n8n Docs (opens in a new tab) · Official docs · accessed
  17. 27.Preemptible compute | Together AI docs (opens in a new tab) · Official docs · accessed
  18. 28.Provisioned throughput | Together AI docs (opens in a new tab) · Official docs · accessed
  19. 29.Support | Together AI docs (opens in a new tab) · Official docs · accessed
  20. 30.Supported models (dedicated model inference) | Together AI docs (opens in a new tab) · Official docs · accessed
  21. 31.Third-party integrations | Together AI docs (opens in a new tab) · Official docs · accessed
  22. 32.Together AI docs index (opens in a new tab) · Official docs · accessed
  23. 33.Together.AI - Dify Marketplace (opens in a new tab) · Official docs · accessed
  24. 34.Together.ai Trust Center (opens in a new tab) · Official docs · accessed

Terms & policies

  1. 35.Privacy and security | Together AI docs (opens in a new tab) · Official docs · accessed
  2. 36.Privacy Policy | Together AI (opens in a new tab) · Official docs · accessed
  3. 37.Terms of Service | Together AI (opens in a new tab) · Official docs · accessed

Status

  1. 38.Together AI status (opens in a new tab) · Official status page · accessed

Secondary sources

  1. 39.Generate AI videos from scripts with DeepSeek, TTS, and Together.ai | n8n workflow template (opens in a new tab) · Secondary source · accessed
  2. 40.Together AI | LinkedIn (opens in a new tab) · Secondary source · accessed
  3. 41.Together AI Continues European Expansion: Infrastructure Now Live and Operational in Sweden (opens in a new tab) · Secondary source · accessed
  4. 42.Unable to find certain credentials such as eleven labs, together.ai etc | n8n Community (opens in a new tab) · Secondary source · accessed

Alternatives to Together AI

Providers that share a hosting category with Together AI, the most relevant first, spread so that every provider is listed on several provider pages. Not a ranking.

  • Modal profileB200 · H200 · +7 more · None in the EU · Per second
    $0.59≈ €0.527 (estimated conversion from $0.59)/GPU-hrChecked modal.com
  • Hyperstack profileB200 · H200 · +4 more · None in the EU · Per minute
    $0.50≈ €0.447 (estimated conversion from $0.50)/GPU-hrChecked hyperstack.cloud
  • Lambda profileB200 · GH200 Grace Hopper Superchip · +5 more · 1 in the EU · Per minute
    $0.79≈ €0.706 (estimated conversion from $0.79)/GPU-hrChecked lambda.ai

Also sharing categories: Hetzner profile

Compare all 4 side by side

Related comparisons and guides

Categories that list Together AI