Menu

Modal GPU prices: H100, RTX PRO 6000, H200, B200 per hour

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

Serverless GPU and CPU containers for AI, billed per second.

GPU cloudInference hostingManaged AI hostingAI agent hosting

Managed platform · Headquarters: New York, United States

From published sources, not hands-on tested

From
$0.59≈ €0.527 (estimated conversion from $0.59)/GPU-hrT4 16GB · Serverless, flexChecked modal.com
GPU families
9
B200 · H200 · RTX PRO 6000 · +6 more · 10 models
Regions
22
None in the EU
Billing
Per second
Serverless, usage-based, monthly

Prices checked 40 sourcesSources (40)How we check

In short

Modal (Modal Labs, Inc.) is a managed serverless compute platform for AI and data workloads. Container images, GPU types and scaling are defined in Python code (JavaScript and Go SDKs in beta), and Modal runs Functions, Sandboxes, Notebooks and model Endpoints on capacity pooled across multiple clouds, scaling to zero when idle.

Read full summary

GPUs from NVIDIA T4 to B300 are billed per second at published list prices (0.59 to 7.10 USD per GPU-hour on 26 Sep 2026), with CPU, memory, storage and, from 1 Oct 2026, network egress billed separately.

From

$0.59≈ €0.527 (estimated conversion from $0.59)/GPU-hrT4 16GB · Serverless, flexChecked modal.com
Visit Modal Opens modal.com in a new tab · Check the final plan, price and terms there before you order.Compare with alternatives to Modal
Setup guides for Modal
Billing increment
Per second
Egress
$0.037≈ €0.033 (estimated conversion from $0.037) per GBSource
Persistent storage
Yes
Managed service
Yes

Pricing

Prices checked 12 pricesHow we check

Billing: per second, serverless, usage-based, monthly. Lowest pay-as-you-go GPU price: $0.59≈ €0.527 (estimated conversion from $0.59) per GPU-hour (T4 16GB, serverless, flex). Lowest plan price: no monthly fee (Serverless endpoint, usage-based); paid plans from $250.00≈ €223.35 (estimated conversion from $250.00) 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 offers10

Modal GPU offers: 10 offers grouped by GPU; prices per GPU-hour, excl. VAT. Region: global for every offer. Availability: available for every offer.

NVIDIA B200 SXM 180GB180 GB VRAM · 1 offer · from $6.25≈ €5.58 (estimated conversion from $6.25) per GPU-hour

Checked modal.com

  • Serverless, flex$6.25≈ €5.58 (estimated conversion from $6.25)/GPU-hrPer second

NVIDIA H200 SXM 141GB141 GB VRAM · 1 offer · from $4.54≈ €4.06 (estimated conversion from $4.54) per GPU-hour

Checked modal.com

  • Serverless, flex$4.54≈ €4.06 (estimated conversion from $4.54)/GPU-hrPer second

NVIDIA RTX PRO 6000 Blackwell 96GB96 GB VRAM · 1 offer · from $3.03≈ €2.71 (estimated conversion from $3.03) per GPU-hour

Checked modal.com

  • Serverless, flex$3.03≈ €2.71 (estimated conversion from $3.03)/GPU-hrPer second

NVIDIA A100 SXM4 80GB80 GB VRAM · 1 offer · from $2.50≈ €2.23 (estimated conversion from $2.50) per GPU-hour

Checked modal.com

  • Serverless, flex$2.50≈ €2.23 (estimated conversion from $2.50)/GPU-hrPer second

NVIDIA H100 SXM 80GB80 GB VRAM · 1 offer · from $3.95≈ €3.53 (estimated conversion from $3.95) per GPU-hour

Checked modal.com

  • Serverless, flex$3.95≈ €3.53 (estimated conversion from $3.95)/GPU-hrPer second

NVIDIA L40S 48GB48 GB VRAM · 1 offer · from $1.95≈ €1.74 (estimated conversion from $1.95) per GPU-hour

Checked modal.com

  • Serverless, flex$1.95≈ €1.74 (estimated conversion from $1.95)/GPU-hrPer second

NVIDIA A100 40GB40 GB VRAM · 1 offer · from $2.10≈ €1.88 (estimated conversion from $2.10) per GPU-hour

Checked modal.com

  • Serverless, flex$2.10≈ €1.88 (estimated conversion from $2.10)/GPU-hrPer second

NVIDIA A10 24GB24 GB VRAM · 1 offer · from $1.10≈ €0.983 (estimated conversion from $1.10) per GPU-hour

Checked modal.com

  • Serverless, flex$1.10≈ €0.983 (estimated conversion from $1.10)/GPU-hrPer second
Show all 10 offers (from $0.59≈ €0.527 (estimated conversion from $0.59) per GPU-hour)

NVIDIA L4 24GB24 GB VRAM · 1 offer · from $0.80≈ €0.715 (estimated conversion from $0.80) per GPU-hour

Checked modal.com

  • Serverless, flex$0.80≈ €0.715 (estimated conversion from $0.80)/GPU-hrPer second

NVIDIA T4 16GB16 GB VRAM · 1 offer · from $0.59≈ €0.527 (estimated conversion from $0.59) per GPU-hour

Checked modal.com

  • Serverless, flex$0.59≈ €0.527 (estimated conversion from $0.59)/GPU-hrPer second

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 offers10
NVIDIA B200 SXM 180GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.001736/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. Modal lists 192 GB HBM3e for its B200 (this site uses NVIDIA's 180 GB HGX B200 figure). gpu='B200+' lets Modal run the request on a B300 while billing the B200 rate. Up to 8 per container.
NVIDIA H200 SXM 141GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.001261/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. SXM variant (docs). Up to 8 per container.
NVIDIA RTX PRO 6000 Blackwell 96GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.000842/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. Modal lists 96 GB GDDR7; which RTX PRO 6000 Blackwell edition (Server or Workstation) is used is not disclosed, so the offer is listed as the edition-neutral RTX PRO 6000 Blackwell 96GB. Multi-GPU limits for this type are not documented.
NVIDIA A100 SXM4 80GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.000694/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. Current docs do not state the form factor; mapped to SXM because Modal's own Oct 2024 A100 pricing article lists its A100 80GB as SXM (medium confidence). Up to 8 per container.
NVIDIA H100 SXM 80GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.001097/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. All H100s on Modal are SXM (docs). gpu='H100' may be automatically upgraded to an H200 at the same price; 'H100!' opts out. Up to 8 per container.
NVIDIA L40S 48GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.000542/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. Up to 8 per container. Modal's docs recommend the L40S as the default starting point for inference.
NVIDIA A100 40GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.000583/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. gpu='A100' may be automatically upgraded to an 80 GB A100 at no extra cost ('A100-40GB' pins 40 GB). Up to 8 per container.
NVIDIA A10 24GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.000306/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. Pricing page label 'Nvidia A10' (24 GB); Modal's pricing data and some doc examples use the type string 'A10G' (current docs list gpu='A10'). Up to 4 per container (96 GB).
NVIDIA L4 24GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.000222/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. Up to 8 per container.
NVIDIA T4 16GB · Serverless, flex
List price per GPU-hour with no region pin (the page shows $0.000164/sec by default; hourly value from the page's per-hour view). GPU only: CPU (0.0473 USD per physical core-hour, minimum 0.125 core) and memory (0.008 USD per GiB-hour) are billed on top, per second. Pinning a region multiplies the whole container price by 1.15x (broad) or 1.75x (narrow). GPU Functions are preemptible and cannot be made non-preemptible. Up to 8 per container.

Compare these prices with other providers’: H100 prices comparedH200 prices comparedL40S prices compared.

Plans3

Serverless endpoint3 plans · no monthly fee (usage-based) · paid from $250.00≈ €223.35 (estimated conversion from $250.00) per month

  • StarterGlobal · ManagedChecked modal.comNo monthly fee
  • TeamGlobal · ManagedChecked modal.com$250.00≈ €223.35 (estimated conversion from $250.00)/month
  • EnterpriseGlobal · ManagedChecked · Stalemodal.comNot verified
Research notes for these plans3
Starter
Platform tier, not a compute bundle. Includes 30 USD of compute per month, up to 3 seats, 100 containers and 10 concurrent GPUs, 200 deployed apps, 1-day log retention, 5 deployed cron jobs, region selection and community Slack support. Compute is billed per second on top of the plan: Functions CPU 0.0473 USD per physical core-hour (1 core = 2 vCPU) and memory 0.008 USD per GiB-hour; Sandboxes and Notebooks CPU 0.1419 USD per core-hour and memory 0.024 USD per GiB-hour; GPUs at the per-GPU rates listed under GPU offers. Included compute does not cover Shared Endpoint tokens (billed per token from the first request since 1 Sep 2026). Egress allowance 1 TiB per billing cycle once egress billing starts on 1 Oct 2026.
Team
Platform fee of 250 USD per month plus usage. Includes 100 USD of compute per month, unlimited seats, 5,000 containers and 50 concurrent GPUs, 1,000 deployed apps, 30-day log retention, unlimited cron jobs, custom domains, static IP proxy (Modal Proxies are labelled Beta in the docs), deployment rollbacks (3 versions), environment-level budgets and RBAC. Compute is billed per second on top of the plan: Functions CPU 0.0473 USD per physical core-hour (1 core = 2 vCPU) and memory 0.008 USD per GiB-hour; Sandboxes and Notebooks CPU 0.1419 USD per core-hour and memory 0.024 USD per GiB-hour; GPUs at the per-GPU rates listed under GPU offers. Included compute does not cover Shared Endpoint tokens (included token usage is 0 on every plan). Egress allowance 10 TiB per billing cycle from 1 Oct 2026.
Enterprise
Custom pricing (contact sales). Adds volume-based discounts, higher GPU concurrency, private Slack support, embedded ML engineering services, audit logs, SAML SSO and HIPAA (BAA), plus use of AWS/GCP marketplace committed spend. Included token usage for Shared Endpoints is 0 on this plan too (compare table). Egress allowance 100 TiB per billing cycle from 1 Oct 2026.

Storage and data transfer

Modal Volumes (distributed file system)
$0.084≈ €0.075 (estimated conversion from $0.084) per GB-month
Network volume · Persistent0.09 USD per GiB-month (about 0.0838 USD per decimal GB-month) with 1 TiB per month free. Pricing page: Modal Plan Pricing (opens in a new tab)
MoreUsage is measured from a daily snapshot of total storage and deleted data can remain billable for up to 4 days (Volumes guide). Durable storage sits in US object storage regardless of compute region (data-residency guide). Volumes v2 (beta) lifts the v1 500,000-file limit and is the only Volume type in scope of the HIPAA BAA. Reads and writes to Volumes do not count as network egress. Published as USD 0.09 per GiB-month; stored per GB (1 GiB = 1.073741824 GB).
Ephemeral container disk (SSD scratch space)
Not verified
Included disk · Not persistentContainers get local SSD scratch space with a default per-container quota of 512 GiB, raisable to 3.0 TiB with ephemeral_disk. Docs: Configuring CPU, memory, and disk (Modal Docs) (opens in a new tab)
MoreDisk requests are billed by raising the memory request to at least one twentieth of the disk request (for example, 500 GiB of disk raises the memory request to 25 GiB), and only if the memory request is not already set higher, so there is no per-GB-month price and a large enough memory request makes the extra disk free. Contents are lost when the container stops.
Cloud bucket mounts (your own S3, R2 or GCS bucket)
Not verified
Object storage · PersistentMounts a customer-owned AWS S3, Cloudflare R2 or Google Cloud Storage bucket (read-write or read-only, built on AWS mountpoint-s3). Docs: Cloud bucket mounts (Modal Docs) (opens in a new tab)
MoreStorage is billed by the bucket provider, not Modal. Modal's data-residency guide recommends this to keep data at rest in a chosen region. Uploads through bucket mounts count as Modal network egress from 1 Oct 2026.
Egress (data out)
$0.037≈ €0.033 (estimated conversion from $0.037) per GB
0.04 USD per GiB (about 0.0373 USD per decimal GB) beyond a workspace-wide allowance per billing cycle: Starter 1 TiB, Team 10 TiB, Enterprise 100 TiB. Docs: Network egress billing (Modal Docs) (opens in a new tab)
MoreEgress charging starts 1 Oct 2026 (usage is shown but not charged in Sep 2026; first bill with egress on 1 Nov 2026). Counts container network egress, private traffic to other Modal containers and Cloud Bucket Mount uploads; Volume reads and writes are excluded. Published as USD 0.04 per GiB; stored per GB (1 GiB = 1.073741824 GB).

Before you compare

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

Suited to
  • Bursty or unpredictable GPU inference that benefits from scale-to-zero and per-second billing
  • Python teams that want containers, GPUs, schedules and scaling defined in code instead of managing VMs or Kubernetes
All 4 uses it suits
Main limitation
GPU list prices exclude CPU (0.0473 USD per physical core-hour for Functions, 0.1419 for Sandboxes and Notebooks) and memory (0.008 USD per GiB-hour for Functions, 0.024 for Sandboxes and Notebooks), which are billed on top
All 8 limitations
Price basis
GPU prices per GPU-hour, billed per second; egress priced per GB; 3 storage options listed
All prices, storage and egress
Alternatives
OVHcloud provider pageDigitalOcean provider pageTogether AI provider pageLambda provider page
Compare side by side with ModalHow they differ

Who it fits

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

Suited to

  • Bursty or unpredictable GPU inference that benefits from scale-to-zero and per-second billing
  • Python teams that want containers, GPUs, schedules and scaling defined in code instead of managing VMs or Kubernetes
  • Running untrusted or agent-generated code in isolated Sandboxes at high concurrency
  • Parallel batch jobs and short-to-medium fine-tuning runs on up to 8 GPUs per container

Good fit for

  • Fine-tuningStrong fitPer-second GPU billing, official fine-tuning examples (Unsloth LLM fine-tuning, Flux LoRA, Whisper) and Volumes for weights fit short-to-medium fine-tuning jobs; checkpointing is advised because GPU Functions are preemptible.
  • Batch inferenceStrong fitWork fans out across many containers (.map, job queues) with per-second billing and scale-to-zero; official vLLM-throughput and batched Whisper examples. Per-plan GPU concurrency limits apply (10 on Starter, 50 on Team).
  • Real-time inferenceStrong fitAutoscaling web Functions, Servers and Dedicated Endpoints scale to zero; minimum containers can be kept warm at normal per-second rates to avoid cold starts. Requests route through us-east by default unless a routing region is set.
  • TrainingSuitableSingle-node training on up to 8 GPUs (A100, H100, H200, B200, B300) with Volumes for checkpoints. Functions are preemptible and each execution is capped at 24 hours, so long runs need checkpoint/resume; multi-node clusters are in private beta.
Show 4 more
  • AI agentsSuitableSandboxes are built for running agent- or model-generated code at high concurrency, with snapshots, tunnels and resource limits, and Hermes Agent documents Modal as a backend. Each Sandbox lives at most 24 hours, so always-on agent gateways need a snapshot-and-restart pattern or another host.
  • RAG applicationsSuitableOfficial RAG and embedding examples (multimodal PDF chat, TEI embeddings) and web serving are available; a vector database must be brought separately or self-run, and Modal Dicts and Queues have retention limits (7 days after last access; 24 h default TTL).
  • RenderingSuitableAn official Blender render-farm example renders frames in parallel across GPU containers billed per second; there are no remote-desktop or GUI sessions.
  • Private LLM hostingSuitableOpen-weight or custom models can be served on your own containers (vLLM/SGLang examples, Dedicated Endpoints with custom weights), and Modal's inference endpoints do not store request or response payloads. The platform is multi-tenant (gVisor isolation), durable storage and logs are in the US, and a HIPAA BAA requires the Enterprise plan.

Look elsewhere for

No workload is rated a partial fit or not suitable.

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

Strengths and limitations

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

Strengths

  • Per-second billing for GPU, CPU and memory with no charge once containers scale to zero (a configurable 60 s idle window is billed by default)
  • One published price list from T4 to B300, including L40S, A100 40/80 GB, RTX PRO 6000, H100 and H200
  • Up to 8 GPUs per container for most GPU types (A10: 4); multi-node clusters in private beta
  • EEA container regions (eu, eu-west, eu-north, eu-south) plus UK, with an eu-west (Dublin) routing region
Show 3 more
  • Per Modal's security pages: SOC 2 Type 2 audited, a published DPA with EU SCCs, and a HIPAA BAA on Enterprise
  • Sandboxes with filesystem, directory and memory snapshots, tunnels, and a beta VM runtime that can run Docker
  • Large set of official, maintained examples (vLLM, SGLang, TensorRT-LLM, LangGraph agents, fine-tuning, Whisper, Flux)

Limitations

  • GPU list prices exclude CPU (0.0473 USD per physical core-hour for Functions, 0.1419 for Sandboxes and Notebooks) and memory (0.008 USD per GiB-hour for Functions, 0.024 for Sandboxes and Notebooks), which are billed on top
  • Pinning to a region costs 1.15x (broad, e.g. eu) or 1.75x (narrow, e.g. eu-west) of base price, and per-region GPU availability is not published
  • Volumes, Images, Dicts, Queues, Secrets, application logs and large or async Function payloads are stored in the US even when compute is pinned to the eu (EEA) regions
  • Network egress is billable from 1 Oct 2026 at 0.04 USD per GiB beyond the plan allowance (1 TiB Starter, 10 TiB Team, 100 TiB Enterprise)
Show 4 more
  • Functions are preemptible by default; GPU Functions cannot be made non-preemptible (GPU Sandboxes are preemptible too), and non-preemptible CPU Functions cost 3x CPU/memory list price
  • Sandboxes live at most 24 hours and Function executions time out after at most 24 hours
  • No public uptime SLA; an SLA is only attached to Service Orders for subscription plans that include one
  • Code-first platform: no VM-style SSH, root access to hosts, bare metal or Kubernetes

Infrastructure

Regions

22 regions across 6 countries; none is recorded as an EU region.

Europe (5)
EU North · eu-north; EU South · eu-south; EU West · eu-west; European Economic Area (broad); United Kingdom · uk
North America (7)
Canada; Mexico; United States (broad); US Central · us-central; US East · us-east; US South · us-south; US West · us-west
South America (1)
South America · sa
Asia (6)
Asia-Pacific (broad) · ap; Asia-Pacific Northeast · ap-northeast; Asia-Pacific South · ap-south; Asia-Pacific Southeast · ap-southeast; Japan; Middle East · me
Oceania (2)
Australia; Melbourne · Australia · ap-melbourne
Africa (1)
Africa

Source: Docs: Region selection (Modal Docs) (opens in a new tab)

Region notes12
United States (broad)
Broad container region: pinning a Function, Sandbox, Server or Dedicated Endpoint here multiplies its price by 1.15x (Functions and Sandboxes per the region-selection guide; Servers and Dedicated Endpoints per the data-residency and Dedicated Endpoints guides). Exact data-center locations are not published; GPU availability per region is not published.
US East
Narrow container region: pinning here multiplies the price by 1.75x. The us-east routing region (Virginia, USA) is Modal's default request-routing region. Exact data-center locations are not published; GPU availability per region is not published.
US Central, US South, Japan, Australia, Canada, Mexico
Narrow container region: pinning here multiplies the price by 1.75x. Exact data-center locations are not published; GPU availability per region is not published.
US West
Narrow container region: pinning here multiplies the price by 1.75x. A us-west (Oregon, USA) routing region is also available. Exact data-center locations are not published; GPU availability per region is not published.
European Economic Area (broad)
Broad container region: pinning a Function, Sandbox, Server or Dedicated Endpoint here multiplies its price by 1.15x (Functions and Sandboxes per the region-selection guide; Servers and Dedicated Endpoints per the data-residency and Dedicated Endpoints guides). Modal describes it as the European Economic Area; country not disclosed. Durable storage (Volumes, Images, Dicts, Queues, Secrets), application logs and large or async Function payloads stay in the US even when compute is pinned here (see data-residency guide). The listing records EU status as not verified because Modal describes this group as the European Economic Area without naming countries (AIHostCompare counts EU member states only).
EU West
Narrow container region: pinning here multiplies the price by 1.75x. Part of the eu (EEA) group; country not disclosed for the container region. The eu-west routing region is Dublin, Ireland. US-resident storage and logs caveat applies. Exact data-center locations are not published; GPU availability per region is not published. The listing records EU status as not verified because Modal describes this group as the European Economic Area without naming countries (AIHostCompare counts EU member states only).
EU North, EU South
Narrow container region: pinning here multiplies the price by 1.75x. Part of the eu (EEA) group; country not disclosed. US-resident storage and logs caveat applies. Exact data-center locations are not published; GPU availability per region is not published. The listing records EU status as not verified because Modal describes this group as the European Economic Area without naming countries (AIHostCompare counts EU member states only).
United Kingdom
Narrow container region: pinning here multiplies the price by 1.75x. Outside the EU/EEA (UK GDPR applies). Exact data-center locations are not published; GPU availability per region is not published.
Asia-Pacific (broad)
Broad container region: pinning a Function, Sandbox, Server or Dedicated Endpoint here multiplies its price by 1.15x (Functions and Sandboxes per the region-selection guide; Servers and Dedicated Endpoints per the data-residency and Dedicated Endpoints guides). Spans the Asia-Pacific group, which also lists Japan and Australia; countries not disclosed.
Asia-Pacific Northeast, Asia-Pacific Southeast, Middle East, South America, Africa
Narrow container region: pinning here multiplies the price by 1.75x. Country not disclosed. Exact data-center locations are not published; GPU availability per region is not published.
Asia-Pacific South
Narrow container region: pinning here multiplies the price by 1.75x. Country not disclosed for the container region; the ap-south routing region is Mumbai, India. Exact data-center locations are not published; GPU availability per region is not published.
Melbourne
Narrow container region: pinning here multiplies the price by 1.75x. Country taken from the region name (Melbourne). Exact data-center locations are not published; GPU availability per region is not published.

Frameworks

How Modal 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 Modal.

LangGraph
Documented setup
Official example: 'Build a coding agent with Modal Sandboxes and LangGraph' (featured as 'Sandbox a LangGraph agent's code'); the agent's generated code runs in a Modal Sandbox. Docs: Featured examples (Modal Docs) (opens in a new tab)
MoreThe example's go() function compiles and runs the LangGraph graph inside a Modal Function (called with go.remote() from a local entrypoint), so LangGraph itself runs on Modal (per the example page as of 27 Sep 2026).
Ollama hosting providers
Documented setup
First-party article 'How to run Ollama' (15 Sep 2024) with code in a third-party repo. Docs: How to run Ollama (Modal article, 15 Sep 2024) (opens in a new tab)
MoreIt uses Image.copy_local_file, whereas the current Images guide documents add_local_file, and it is not in the current examples index, so expect to adapt it.
PyTorch
Documented setup
The GPU guide's quickstart installs torch; official examples cover torch.compile serving, PyTorch profiling and training runs. Docs: GPU acceleration (Modal Docs) (opens in a new tab)
SGLang
Documented setup
Official examples: 'Low latency Qwen 3.6 with SGLang and Modal' (sglang_low_latency, on the featured examples page) and 'Serve Inkling-Small on Modal with SGLang' (inkling_small); the llm-serving directory also has sglang_snapshot and sglang_vlm. Docs: Featured examples (Modal Docs) (opens in a new tab)
vLLM hosting providers
Documented setup
Official, maintained examples: 'Run OpenAI-compatible LLM inference with Gemma and vLLM' (llm_inference, featured as 'Deploy an OpenAI-compatible LLM service') and 'Run LLM inference at maximum throughput' (vllm_throughput); source in modal-labs/modal-examples (vllm_inference.py last changed 9 Jul 2026). Docs: Featured examples (Modal Docs) (opens in a new tab)
Text Generation Inference
Manual setup
No current official TGI example (the llm-serving examples cover vLLM, SGLang and TensorRT-LLM). Modal can run registry container images with GPUs, so TGI would be a manual, untested setup. Docs: modal-labs/modal-examples: llm-serving directory (opens in a new tab)
Dify
Not verified
No official or community guide for running Dify on Modal was found (as of 26 Sep 2026).
Flowise
Not verified
No official or community guide for running Flowise on Modal was found (as of 26 Sep 2026).

Model API

Modal serves the models the framework calls; the framework itself runs elsewhere.

OpenClaw
Documented setup
Model API only: Modal's blog (Feb 2026, updated Apr 2026) and its modal-projects/modal-jazz repo document pointing a separately run OpenClaw at a Modal-hosted, OpenAI-compatible model endpoint. Docs: Try GLM-5.1 on Modal (Modal blog, 11 Feb 2026, updated 7 Apr 2026) (opens in a new tab)
MoreThere is no guide for hosting the OpenClaw gateway on Modal; running it inside Modal (for example in a Node.js Sandbox) would be a manual, untested setup limited by the 24-hour Sandbox lifetime. Shown as Documented setup because Modal documents this in its own blog and repository.

Integration

Modal is a sandbox, backend or connector for the framework; the framework itself runs elsewhere.

Hermes Agent
Documented setup
Hermes Agent's own docs list Modal as a terminal backend (terminal.backend: modal, MODAL_TOKEN_ID/MODAL_TOKEN_SECRET): the agent's shell commands run in Modal sandboxes with configurable CPU, memory and disk and optional filesystem-snapshot persistence, while the Hermes process runs on your host. Secondary source: Hermes Agent Configuration (Nous Research docs) (opens in a new tab)
n8n
Community guide
Integration only, from community evidence: a third-party blog post (26 Feb 2026) uses the pattern 'n8n (Trigger) -> Modal (AI Processing) -> n8n (Output/Routing)', with n8n running elsewhere and calling Modal-deployed code. Secondary source: Why n8n Is Still My Secret Weapon for AI Agent Workflows in 2026 (Tarek Alaaddin, 26 Feb 2026) (opens in a new tab)
MoreNo guide for hosting n8n itself on Modal was found (as of 26 Sep 2026). Shown as Community guide rather than Not verified because that third-party post documents the integration.

Modal is not listed on the OpenClaw hosting and Hermes Agent 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
Images can be defined from an existing Dockerfile or pulled from a container registry. Running a Docker daemon inside a container is supported in VM Sandboxes (beta, CPU-only). Docs: Images (Modal Docs) (opens in a new tab)
Root access
Partial
Code runs as root inside gVisor-sandboxed containers (the modal shell prompt is root); VM Sandboxes (beta, CPU-only) add a real Linux kernel with custom init, FUSE and eBPF. Docs: VM Sandboxes (Modal Docs, beta) (opens in a new tab)
MoreNo host or bare-metal root.
SSH access
Partial
No built-in SSH endpoint. modal shell opens an interactive shell in a container, and an SSH server can be exposed through an unencrypted TCP tunnel (tunnels carry no extra charge). Docs: Tunnels (Modal Docs) (opens in a new tab)
API access
Yes
Python SDK and CLI (primary); JavaScript/TypeScript and Go SDKs are labelled Beta in the docs navigation. Docs: Introduction (Modal Docs) (opens in a new tab)
MoreEndpoints expose OpenAI-compatible (Chat Completions, Responses, Embeddings) and Anthropic-compatible (Messages) APIs per the Endpoints guide (source Endpoints (Modal Docs) (in Sources)).
Jupyter notebooks
Yes
Modal Notebooks are hosted, collaborative Jupyter notebooks with GPU kernels (up to 8 A100s or H100s), billed at Sandbox CPU/memory rates while the kernel runs. Docs: Modal Notebooks (Modal Docs) (opens in a new tab)
Templates
Yes
Library of official, runnable examples (LLM serving, fine-tuning, audio, image/video, sandboxed agents) and one-command Dedicated Endpoints for Modal Library models. Docs: Featured examples (Modal Docs) (opens in a new tab)
MoreNo one-click templates for apps such as n8n or Dify were found.
Multi-GPU instances
Yes
Up to 8 GPUs per container on one machine for B300, B200, H200, H100, A100, L4, T4 and L40S (A10: up to 4); multi-node training is in private beta. Docs: GPU acceleration (Modal Docs) (opens in a new tab)
Bare metal
No
Workloads run in gVisor-sandboxed containers (or beta VM Sandboxes); Modal's DPA states it does not operate physical servers. Docs: Security and privacy at Modal (Modal Docs) (opens in a new tab)
Kubernetes
No
No Kubernetes service or kubectl access; the docs present Modal as a way to avoid managing Kubernetes. Docs: Introduction (Modal Docs) (opens in a new tab)
Show all 21 features

Management

Managed service
Yes
Modal hosts and schedules all containers on capacity pooled across multiple clouds; there are no servers or clusters to operate. Docs: Introduction (Modal Docs) (opens in a new tab)
Serverless
Yes
Containers scale to zero; billing covers load time, execution and a configurable idle window (60 s by default). Pricing page: Modal Plan Pricing (opens in a new tab)

Storage and data

Persistent storage
Yes
Modal Volumes (0.09 USD per GiB-month, 1 TiB free) and cloud bucket mounts for customer-owned S3, R2 or GCS buckets. Docs: Volumes (Modal Docs) (opens in a new tab)
Snapshots
Partial
Sandboxes support filesystem and directory snapshots (30-day default retention, configurable or indefinite) and experimental memory snapshots (7 days, not with GPUs). Docs: Sandbox snapshots (Modal Docs) (opens in a new tab)
MoreSnapshots exclude mounted Volumes and are stored in the US. Functions can use memory snapshots to speed up cold starts.
Managed backups
No
No user-facing backup or restore service for Volumes is documented; the shared responsibility model makes customers responsible for their own backups (the DPA says Modal backs up customer data internally at least daily). Docs: Security and privacy at Modal (Modal Docs) (opens in a new tab)

Billing

Per-second billing
Yes
GPU, CPU and memory are all priced per second. Pricing page: Modal Plan Pricing (opens in a new tab)
Spot pricing
No
No discounted spot tier. Docs: Preemption (Modal Docs) (opens in a new tab)
MoreFunctions are preemptible by default at list price; non-preemptible CPU Functions cost 3x the CPU/memory list price and GPU Functions cannot be made non-preemptible. GPU Sandboxes are also preemptible; CPU Sandboxes are not.
Reserved pricing
Partial
No published reserved-capacity prices. Enterprise lists volume-based discounts, and AWS/GCP marketplace committed spend can be applied (contact sales). Pricing page: Modal Plan Pricing (opens in a new tab)

Regions and compliance

EU region
Not verified
Modal names no country for eu, eu-west, eu-north or eu-south and describes the eu group as the European Economic Area, so a region in an EU member state is not verified (the EEA also includes Iceland, Liechtenstein and Norway). Docs: Region selection (Modal Docs) (opens in a new tab)
MoreContainers can be pinned to eu (EEA), eu-west, eu-north or eu-south (or uk) at a 1.15x or 1.75x price multiplier, with an eu-west (Dublin) routing region. Durable storage (Volumes, Images, Dicts, Queues, Secrets), application logs, snapshots and large or async Function payloads are stored in the US (data-residency guide).
SOC 2 report
Yes
SOC 2 Type 2 audit completed; report available via trust.modal.com. The pricing page lists SOC 2 on all plans. Docs: Security and privacy at Modal (Modal Docs) (opens in a new tab)
GDPR data processing agreement
Yes
Published Data Processing Addendum (effective 31 Oct 2025) covering EU GDPR, UK GDPR and Swiss law with EU SCCs and the UK IDTA. Docs: Modal Labs, Inc. Data Processing Addendum (effective 31 Oct 2025) (opens in a new tab)
MoreModal's data-residency guide says customers must configure regions and storage to meet GDPR obligations.
Customer-managed keys (BYOK)
Partial
Customer Supplied Encryption Keys are in alpha and currently cover only Sandbox Directory Snapshots (experimental SDK parameter); Modal does not persist the key. Docs: Customer Supplied Encryption Keys (Modal Docs) (opens in a new tab)
MoreThird-party API keys can be stored as Modal Secrets.

SLA and support

Uptime SLA
None published
No public uptime commitment. Docs: Modal Software as a Service Agreement (effective May 2026) (opens in a new tab)
MoreThe Terms (effective May 2026) say an SLA is attached to the Service Order only when the customer buys a subscription plan that includes uptime and support SLAs. The status page shows historical 90-day component uptime (e.g. Functions 100%, Sandboxes 99.989% on 26 Sep 2026), which is not a commitment.
Support channels
4 listed
  • Community Slack (all plans)
  • Email (support@modal.com)
  • Private Slack (Enterprise)
  • Embedded ML engineering services (Enterprise)
Channels per plan from the pricing comparison table; support@modal.com is listed on the company page. No published response-time targets were found. Pricing page: Modal Plan Pricing (opens in a new tab)

Sources and verification

40 sources, accessed between 26 Sep 2026 and 20 Oct 2026. Official provider pages are listed first; the 3 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 Modal 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, 18:42 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, plan limit and region code in this listing against Modal's live pricing page, documentation and status page. Modal's current docs do not state the A100 80GB form factor (listed as SXM, as in an earlier Modal article) or its RTX PRO 6000 edition. Not verified: the countries of most container regions, GPU availability per region and Shared Endpoint token prices. Later corrections are listed on the methodology page.

Commercial relationship

AIHostCompare has no active affiliate link with Modal; 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.Modal Plan Pricing (opens in a new tab) · Official pricing page · accessed
  2. 2.Modal pricing page code (FAQ answers, CPU and memory hourly rates) (opens in a new tab) · Official pricing page · accessed
  3. 3.Modal pricing page GPU price table (data behind the per-hour toggle) (opens in a new tab) · Official pricing page · accessed
  4. 4.Modal pricing page Sandbox and Notebook compute rates (opens in a new tab) · Official pricing page · accessed

Documentation

  1. 5.About Modal (company page) (opens in a new tab) · Official docs · accessed
  2. 6.Cloud bucket mounts (Modal Docs) (opens in a new tab) · Official docs · accessed
  3. 7.Configuring CPU, memory, and disk (Modal Docs) (opens in a new tab) · Official docs · accessed
  4. 8.Customer Supplied Encryption Keys (Modal Docs) (opens in a new tab) · Official docs · accessed
  5. 9.Data residency on Modal (Modal Docs) (opens in a new tab) · Official docs · accessed
  6. 10.Dedicated Endpoints (Modal Docs) (opens in a new tab) · Official docs · accessed
Show all 40 sources

Documentation

  1. 11.Endpoints (Modal Docs) (opens in a new tab) · Official docs · accessed
  2. 12.Featured examples (Modal Docs) (opens in a new tab) · Official docs · accessed
  3. 13.GPU acceleration (Modal Docs) (opens in a new tab) · Official docs · accessed
  4. 14.How much is an Nvidia A100? (Modal article, 31 Oct 2024) (opens in a new tab) · Official docs · accessed
  5. 15.How to run Ollama (Modal article, 15 Sep 2024) (opens in a new tab) · Official docs · accessed
  6. 16.Images (Modal Docs) (opens in a new tab) · Official docs · accessed
  7. 17.Introduction (Modal Docs) (opens in a new tab) · Official docs · accessed
  8. 18.Modal Changelog (opens in a new tab) · Official docs · accessed
  9. 19.Modal Labs, Inc. Data Processing Addendum (effective 31 Oct 2025) (opens in a new tab) · Official docs · accessed
  10. 20.Modal Notebooks (Modal Docs) (opens in a new tab) · Official docs · accessed
  11. 21.modal-labs/modal-examples: llm-serving directory (opens in a new tab) · Official docs · accessed
  12. 22.Network egress billing (Modal Docs) (opens in a new tab) · Official docs · accessed
  13. 23.Partner with Modal (opens in a new tab) · Official docs · accessed
  14. 24.Preemption (Modal Docs) (opens in a new tab) · Official docs · accessed
  15. 25.Region selection (Modal Docs) (opens in a new tab) · Official docs · accessed
  16. 26.Sandbox resources and pricing (Modal Docs) (opens in a new tab) · Official docs · accessed
  17. 27.Sandbox snapshots (Modal Docs) (opens in a new tab) · Official docs · accessed
  18. 28.Sandboxes (Modal Docs) (opens in a new tab) · Official docs · accessed
  19. 29.Shared Endpoints (Modal Docs) (opens in a new tab) · Official docs · accessed
  20. 30.Timeouts (Modal Docs) (opens in a new tab) · Official docs · accessed
  21. 31.Try GLM-5.1 on Modal (Modal blog, 11 Feb 2026, updated 7 Apr 2026) (opens in a new tab) · Official docs · accessed
  22. 32.Tunnels (Modal Docs) (opens in a new tab) · Official docs · accessed
  23. 33.VM Sandboxes (Modal Docs, beta) (opens in a new tab) · Official docs · accessed
  24. 34.Volumes (Modal Docs) (opens in a new tab) · Official docs · accessed

Terms & policies

  1. 35.Modal Software as a Service Agreement (effective May 2026) (opens in a new tab) · Official docs · accessed
  2. 36.Security and privacy at Modal (Modal Docs) (opens in a new tab) · Official docs · accessed

Status

  1. 37.Modal Labs status page (opens in a new tab) · Official status page · accessed

Secondary sources

  1. 38.Hermes Agent Configuration (Nous Research docs) (opens in a new tab) · Secondary source · accessed
  2. 39.Modal company page on LinkedIn (opens in a new tab) · Secondary source · accessed
  3. 40.Why n8n Is Still My Secret Weapon for AI Agent Workflows in 2026 (Tarek Alaaddin, 26 Feb 2026) (opens in a new tab) · Secondary source · accessed

Alternatives to Modal

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

Also sharing categories: Lambda profile

Compare all 4 side by side

Related comparisons and guides

Categories that list Modal