What it is
Modal lets you run Python code on cloud CPUs and GPUs without managing servers. You add a few lines to your code to say what hardware and packages a function needs, and Modal starts containers when requests arrive and shuts them down when they stop. Teams use it to serve their own models, run batch jobs, fine-tune, and host code sandboxes for AI agents.
The Starter plan is free and includes $30 of compute each month for up to 3 people. Team is $250 a month plus usage, with $100 of included compute. GPUs are billed per second; an H100 works out to about $3.95 an hour, an A100 80GB to about $2.50 and a B200 to about $6.25. Modal also sells shared inference endpoints billed per token.
Our review
Modal is for people who want to run their own code, not call someone else's model. You write ordinary Python, decorate the functions that need a GPU, and Modal builds the container, scales it and tears it down. Against renting servers on AWS or Google Cloud there is no cluster to run and nothing billed once containers scale to zero.
The headline rates are competitive and the multipliers are where bills grow. The pricing page works out to about $3.95 an hour for an H100 and $2.50 for an A100 80GB. Pinning a region costs 1.15x to 1.75x, non-preemptible execution 3x, and sandboxes and notebooks are always billed at the 3x rate. Containers stay up for 60 seconds after the last request by default, and that idle time is billed too.
Cold starts are the engineering trade-off. Modal's docs say containers boot in about a second, but loading a model of tens of gigabytes can take minutes unless weights are cached or memory snapshots are used. Keeping min_containers warm fixes latency by paying for idle GPUs, which erodes the per-second argument for steady traffic.
The plan ladder has a gap. Starter allows 10 GPUs at once and keeps logs for one day; Team at $250 a month raises that to 50 GPUs and 30 days of logs. Audit logs, Okta SSO and HIPAA are Enterprise-only, though Modal holds a SOC 2 Type II report for all customers and runs each container in a gVisor or microVM sandbox. Function inputs and outputs are kept for up to 7 days.
For a team that only needs a popular open model behind an API, Together AI or Groq bills by the token and has no containers to tune. Modal's advantage starts when the model, pipeline or agent sandbox is your own.
Where it earns its place
A two-person ML startup serving a fine-tuned 8B model
Put a custom model behind an endpoint without hiring DevOps
Deploy a web endpoint on an L40S at about $1.95 an hour of active time. Bursty traffic suits scale-to-zero; a model that must answer in under a second needs a warm container, which at 720 hours costs about $1,400 a month.
A data engineer embedding 50 million documents once a quarter
Run a large batch job without keeping GPUs reserved
Fan the work out with .map() across many A10s at $1.10 an hour each; Starter's cap of 10 concurrent GPUs makes this slow, so the job usually justifies Team's 50-GPU limit for $250 a month.
An AI agent company running untrusted user code
Give each agent session an isolated sandbox
Modal Sandboxes isolate each session with gVisor or microVMs but bill CPU at 3x the function rate, about $0.142 per core-hour. Inputs and outputs are kept for up to 7 days, which matters for customer data.
Before you commit
- Setup
- pip install modal, run modal setup to authenticate, and modal run on a decorated function; first GPU output takes minutes, not a cluster build.
- Integrations
- Python SDK, web endpoints, cron jobs, volumes, secrets, and AWS or GCP committed-spend billing through cloud marketplaces on Enterprise.
- Data and privacy
- SOC 2 Type II; function inputs and outputs kept up to 7 days, endpoint payloads not stored; HIPAA only on Enterprise; logs kept 1 day on Starter and 30 on Team.
- Cost at scale
- One H100 running around the clock is about $2,840 a month at base rates, rising to about $8,530 if it must be non-preemptible, plus the $250 Team fee.
- Limits
- Starter allows 100 containers and 10 GPUs at once, Team 5,000 and 50; idle containers are billed for 60 seconds by default; shared endpoints are billed per token outside included credits.
What's good
- Plain Python decorators instead of Dockerfiles and clusters
- Per-second billing that drops to zero when idle
- H100 about $3.95 an hour, A100 80GB about $2.50
- SOC 2 Type II and gVisor or microVM isolation per container
What's not
- Region pinning, non-preemptible runs and sandboxes cost 1.15x to 3x
- Large models cold-start in minutes without snapshots or warm containers
- $250 a month to move beyond 10 concurrent GPUs
- SSO, audit logs and HIPAA only on Enterprise
Best for
Python developers and ML teams who want to serve custom models or run GPU jobs without managing servers.
Alternatives
Modal: common questions
- Is Modal worth it?
- Modal turns a Python function into a GPU service with less ceremony than any cloud we have read about, and charges a premium the moment you need guarantees. We rate it 4.5 out of 5.
- How much does Modal cost?
- Starter $0 + compute ($30/month credit, 3 seats, 10 GPUs at once); Team $250/month + compute ($100 credit, unlimited seats, 50 GPUs); Enterprise custom. Per second: H100 about $3.95/hour, A100 80GB about $2.50/hour, B200 about $6.25/hour.
- What are the downsides of Modal?
- Region pinning, non-preemptible runs and sandboxes cost 1.15x to 3x; large models cold-start in minutes without snapshots or warm containers; $250 a month to move beyond 10 concurrent GPUs; sSO, audit logs and HIPAA only on Enterprise.
- Who is Modal best for?
- Python developers and ML teams who want to serve custom models or run GPU jobs without managing servers.
- How much does Modal cost for a team or in production?
- One H100 running around the clock is about $2,840 a month at base rates, rising to about $8,530 if it must be non-preemptible, plus the $250 Team fee.
- Is Modal safe for company data?
- SOC 2 Type II; function inputs and outputs kept up to 7 days, endpoint payloads not stored; HIPAA only on Enterprise; logs kept 1 day on Starter and 30 on Team.
- What are the best Modal alternatives?
- The closest alternatives we have reviewed are Replicate, Together AI, Hugging Face and Fireworks AI.
Answers checked Oct 22, 2026.
Make Modal? Show this rating on your site
