Modal review
- Made by
- Modal Labs (CEO Erik Bernhardsson, CTO Akshat Bubna)
- Offices
- New York, Stockholm and San Francisco
- Free plan
- Starter: $0/month with $30 of free compute every month
- Paid plans
- Team $250/month + usage ($100 credit); Enterprise custom
- H100 price
- $0.001097 per second (about $3.95 an hour), GPU only
- Billing
- Per second for GPU, CPU and memory; nothing while scaled to zero
- Funding
- $355M at a $4.65B valuation (May 2026)
- Security
- SOC 2; HIPAA, audit logs and SAML SSO on Enterprise
Modal is a serverless cloud for AI and data work that you control from Python. You add a decorator to a function and say which GPU it needs; Modal builds the container, starts it in the cloud in seconds and shuts it down when the work ends. You pay per second for GPU, CPU and memory, and nothing once it scales to zero. That suits inference that arrives in bursts, batch jobs, evaluations, short fine-tunes and sandboxes for AI-generated code.
An H100 costs $0.001097 a second (about $3.95 an hour), and the free Starter plan includes $30 of compute every month. Modal is ranked in our best GPU cloud providers, where it scores highest of all for developer experience.
Modal is worth it for Python teams whose GPU use is bursty: APIs with uneven traffic, batch jobs, evaluation runs and short fine-tunes. Per-second billing and scale to zero mean you pay only while code runs, and the $30 monthly credit covers real experiments. It is a poor fit for GPUs that must run around the clock, where a rented pod or reserved cluster costs less, and for long multi-node training, because GPU functions are always preemptible and multi-node clusters are still in private beta.
Score breakdown
Best for
- Python developers who want GPUs without managing servers or Kubernetes
- Inference APIs with spiky or unpredictable traffic
- Batch jobs, evaluations and short fine-tunes
- Running AI-generated or untrusted code in isolated Sandboxes
Skip it if
- Always-on GPU workloads, where renting a GPU costs less
- Long multi-node training runs that must not be interrupted
- Teams that want a plain VM with full control of the machine
What we like
- GPUs from a Python decorator, with containers built for you
- Per-second billing and scale to zero
- $30 of free compute every month on the Starter plan
- Wide GPU choice from T4 to B300, up to 8 GPUs per container
- Fast cold starts, including GPU memory snapshots for large models
Watch out for
- GPU functions are always preemptible, so long jobs need checkpoints
- CPU and memory are billed on top of the GPU rate
- Pinning a region costs 1.15 to 1.75 times base prices
- Costs more than a rented GPU for 24/7 workloads
- Multi-node training is still in private beta
Key features, explained
Functions from Python
You write normal Python, add a Modal decorator and define the container image in code: the Python version, pip packages and any system libraries. modal run executes it in the cloud and modal deploy keeps it live as a web endpoint, a scheduled job or a function other code can call. There are no YAML files, Dockerfiles or clusters to manage. Starter allows 5 deployed cron jobs; Team has no limit.
GPU choice
Modal offers T4, L4, A10, L40S, A100 (40 and 80 GB), RTX PRO 6000, H100, H200, B200 and B300. Most types allow up to 8 GPUs per container (A10 up to 4). You can list fallback GPU types in order of preference. Modal may upgrade an H100 request to an H200 at no extra cost unless you ask for H100!, and B200+ lets you get a B200 or B300 billed at the B200 rate.
Autoscaling and cold starts
Functions scale from zero to many containers as requests arrive, then back down. By default a container waits 60 seconds after its last request before shutting down, and that idle time is billed. In a May 2026 post, Modal said GPU memory snapshots cut average vLLM start-up in its tests from about 96 seconds to about 14. You can keep containers warm with min_containers, at the cost of paying for them.
Sandboxes and Notebooks
Sandboxes are isolated containers you create on the fly, for example to run code written by an AI agent safely. They are not preempted unless you attach a GPU. Sandbox and Notebook CPU costs $0.00003942 per core per second and memory $0.00000667 per GiB per second, three times the standard function rates, with GPUs at standard prices. Notebooks give you hosted Jupyter-style notebooks on the same compute.
Storage and building blocks
Volumes provide persistent storage for model weights and datasets at $0.09 per GiB a month, with the first 1 TiB each month free. Secrets store API keys, and distributed Dicts and Queues, included on every plan, pass data between functions. Together they cover what most inference and batch pipelines need without setting up a separate database or bucket.
Training and multi-node clusters
Single-node training on up to 8 GPUs works like any other function. For larger jobs, Modal's @clustered feature runs co-located containers across several hosts, with InfiniBand networking of up to 3,200 Gbps, but multi-node training is in private beta and needs a request to Modal support. Because GPU functions can be preempted, save checkpoints regularly on any long run.
Where Modal ranks
Our scores come from our category rankings, where each tool is scored 0–10 on weighted criteria against its rivals.
| Ranking | Listed as | Position | Score |
|---|---|---|---|
| Best GPU Cloud Providers for AI | Modal | #5 of 10 | 7.8 |
What users say about Modal
Modal's public reviews are few but very positive. Product Hunt shows 5.0 out of 5 from 59 reviews, mostly from founders and engineers running AI products in production. They praise how simple it makes launching GPUs and sandboxes, fast container start-up, the Python-first workflow, and the ability for a small team to ship without building infrastructure. Criticism is rare on review sites, so the caveats come mainly from Modal's own documentation and independent price comparisons: per hour, a Modal H100 costs more than renting one from RunPod or Nebius, CPU and memory are billed separately, GPU functions can be preempted, and large models still face cold starts on the first request.
Users praise
- GPUs and sandboxes that are simple to launch
- Fast spin-up and scale to zero
- Clean, Python-first developer workflow
- Lets small teams ship without infrastructure work
Users complain about
- Higher hourly price than rented GPUs for steady load
- Separate CPU and memory charges make bills harder to estimate
- GPU functions can be preempted
- Cold starts for large models on first request
| Review site | Rating | Reviews | Checked |
|---|---|---|---|
| Product Hunt | 5/5 | 59 | 2026-09 |
Ratings are third-party aggregates and change over time. They are context, not part of our score.
Modal charges per second for the GPU, CPU and memory your code uses, and nothing while it is scaled to zero. An H100 costs $0.001097 a second (about $3.95 an hour), an A100 80 GB about $2.50 and an L4 about $0.80; CPU costs $0.0000131 per core per second and memory $0.00000222 per GiB per second. The Starter plan is free and includes $30 of compute every month; Team costs $250 a month plus usage, with $100 of monthly credit and higher limits; Enterprise is custom.
Prices are Modal's list prices in US dollars, checked on 25 September 2026. Pinning a region costs 1.15 to 1.75 times base prices.
| Plan | Price |
|---|---|
| Starter | $0/month + compute |
| Team | $250/month + compute |
| Enterprise | Custom |
| GPU compute (all plans) | From $0.000164/sec (T4, about $0.59/hour); H100 $0.001097/sec (about $3.95/hour) |
| CPU and memory (all plans) | CPU $0.0000131/core/sec; memory $0.00000222/GiB/sec |
| Sandboxes and Notebooks | CPU $0.00003942/core/sec; memory $0.00000667/GiB/sec; GPUs at standard rates |
| Volumes (storage) | $0.09/GiB/month |
| Startup and academic credits | Free credits (by application) |
Full Modal pricing breakdown, hidden costs and money-saving tips
How Modal billing works
Modal bills three meters per second while a container runs: GPU (for example $0.001097 a second for an H100), CPU ($0.0000131 per physical core per second, minimum 0.125 cores) and memory ($0.00000222 per GiB per second). For CPU and memory you pay for whichever is higher, what you requested or what you used.
Your plan's monthly credit ($30 on Starter, $100 on Team) is used first. Choosing a specific region multiplies prices by 1.15 to 1.75, and forcing CPU functions to be non-preemptible triples CPU and memory prices. Full rates and examples are on our Modal pricing page.
Modal vs renting a GPU: when serverless pays off
Serverless wins when your GPU sits idle a lot. A Modal H100 costs about $3.95 an hour of busy time, while a RunPod Secure Cloud H100 pod costs $3.49 an hour whether it is busy or not. Ignoring CPU and memory, Modal is cheaper as long as the GPU is busy less than about 88% of the time; against RunPod's $2.69 Community Cloud price the break-even is about 68%.
For an API busy two hours a day, that makes Modal far cheaper. For a model serving traffic around the clock, or a multi-day training run, a rented GPU or reserved cluster usually costs less. See RunPod and our Modal alternatives.
Recent changes
- May 2026: Modal published GPU memory snapshots, which it says cut cold starts for large inference servers by several times.
- May 2026: Modal raised $355 million at a $4.65 billion valuation, led by Redpoint Ventures and General Catalyst, up from $1.1 billion in September 2025. SiliconANGLE reported annual revenue of about $300 million.
- GPU range: the price list now runs from T4 to B300, with a
B200+option that can land on either Blackwell GPU at the B200 price.
- Set explicit CPU and memory requests. You pay for the higher of your request or actual use, so oversized requests cost money on every second of GPU time.
- Keep model weights in a Volume or bake them into the image instead of downloading them on every cold start; you pay for the download time.
- Lower the idle window (
scaledown_window) for rarely used endpoints so containers stop billing sooner after the last request. - Allow GPU fallbacks (for example H100 then A100) so jobs start faster when your first choice is busy.
- Avoid region pinning unless you need it for data rules or latency: it adds 15% to 75% to every price.
Modal alternatives
Jargon explained
- Serverless
- A model where the platform starts computers only when your code needs them and stops them afterwards, so you pay for use rather than for idle machines.
- Cold start
- The delay while a new container starts and loads a model before it can answer its first request.
- Preemption
- When the platform stops your running job to reclaim the machine and restarts it elsewhere. Saving checkpoints lets a job resume without losing work.
- Decorator
- A line starting with @ placed above a Python function that changes how it runs. On Modal it tells the platform to run that function in the cloud.
- Scale to zero
- Shutting down every container when there is no work, so the bill drops to nothing between requests.
Frequently asked questions
What is Modal used for?
Modal runs Python code on cloud GPUs and CPUs without you managing servers. Common uses are model inference APIs, batch processing, fine-tuning, evaluation runs and sandboxes for running AI-generated code.
Is Modal free?
The Starter plan costs $0 a month and includes $30 of compute every month, with up to 3 seats and 10 GPUs at once. You pay only for usage above the credit.
How much does an H100 cost on Modal?
As of 25 September 2026, $0.001097 per second, about $3.95 an hour, for the GPU alone. CPU and memory are billed on top. See Modal pricing.
Can Modal functions be interrupted?
Yes. Modal's docs say all functions are preemptible by default, and GPU functions cannot opt out, so long GPU jobs should save checkpoints. Sandboxes without GPUs are not preempted.
Is Modal good for training models?
For fine-tunes and training on up to 8 GPUs, yes. Multi-node training is in private beta, and for weeks-long runs on large clusters a provider such as CoreWeave or Nebius is a better fit.
Does Modal work with languages other than Python?
Modal is built around its Python SDK: you define functions, images and endpoints in Python. The code inside a container can run any program you install in the image.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Modal pricing (Modal)
- Preemption (Modal docs) (Modal)
- Reserving CPU and memory (Modal docs) (Modal)
- Cold start performance (Modal docs) (Modal)
- GPU acceleration (Modal docs) (Modal)
- Multi-node clusters (Modal docs) (Modal)
- How we achieved truly serverless GPUs (Modal)
- About Modal (Modal)
- Serverless AI infrastructure startup Modal Labs seals $355M funding round (SiliconANGLE)
- Modal reviews (Product Hunt)
- ClusterMAX 3.0: the industry standard GPU cloud rating system (SemiAnalysis)
- Runpod GPU cloud pricing (Runpod)
- Replicate pricing (Replicate)
- Together AI pricing (Together AI)
- Lambda pricing (Lambda)
- Nebius AI Cloud prices (Nebius)
- CoreWeave cloud pricing (CoreWeave)