Review

Modal Review 2026: Pricing, Pros, Cons and Alternatives

Modal's per-second GPU, CPU and memory rates, the Starter and Team plans, one inference workload costed, and how Replicate and RunPod price the same job.

Product links may be affiliate links. How we rate 3 alternatives covered
TL;DR

Modal is worth it in 2026 for Python teams with spiky GPU workloads, because it bills per second and scales to zero, so idle time costs nothing once containers shut down. Starter is $0 a month plus compute with $30 of free credit; an L40S GPU costs $0.000542 a second, about $1.95 an hour.

For a steady 24/7 workload it is not the cheapest: a RunPod L40S pod lists $1.09 an hour. Watch the multipliers too: choosing a region raises prices 1.15 to 1.75 times, and non-preemptible execution, which Sandboxes use, costs 3 times the list rate.

Key factsBest for: Python teams running bursty GPU inference, batch jobs…
  • Updated: November 15, 2026
  • Best for: Python teams running bursty GPU inference, batch jobs, fine-tuning or agent Sandboxes.
  • Price as of November 15, 2026: Free plan · Team $250/mo · custom-priced plans too
  • Serverless GPU and CPU platform billed per second, defined in Python code and scaled to zero.
  • Billing: Per second for GPU, CPU and memory, including load time and a 60-second default keep-alive.
  • Free credit: Starter includes $30 of compute a month and Team includes $100.
  • Team plan: Team costs $250 a month plus compute, with unlimited seats and 50 concurrent GPUs.
  • Multipliers: Region selection costs 1.15 to 1.75 times base prices and non-preemptible execution 3 times.

Pros

  • Per-second billing and scale to zero, so idle time after the keep-alive costs nothing.
  • Monthly free compute credit on both Starter and Team.
  • GPUs from T4 to B300 on one rate card.

Cons

  • At full utilization an L40S costs $1,404.86 a month in GPU time, more than an hourly pod.
  • Region selection adds 1.15 to 1.75 times and non-preemptible runs 3 times the list price.
BillingPer second for GPU, CPU and memory, including load time and a 60-second default keep-alive.
Free creditStarter includes $30 of compute a month and Team includes $100.
Team planTeam costs $250 a month plus compute, with unlimited seats and 50 concurrent GPUs.
MultipliersRegion selection costs 1.15 to 1.75 times base prices and non-preemptible execution 3 times.

Modal is the serverless GPU platform developers reach for when they want to run a model from a Python function without managing machines. You decorate a function, pick a GPU, and Modal builds the container, runs it and scales it back to zero. The pricing question is whether paying by the second beats renting a GPU by the hour for your traffic.

Toolradar data: Toolradar, the software directory we run, tracks 44 GPU cloud tools as of November 3, 2026, and 41% of them are free, freemium or open source (Toolradar GPU cloud category). Modal sits in the freemium group through its monthly credit, which makes it cheap to try and usage-priced from the first paid second.

This review covers the rate card, the plan limits, one inference workload costed line by line, and how Replicate and RunPod price the same job. Prices were read on each vendor's pricing page on November 4, 2026.

What is Modal?

Modal is a serverless compute platform for AI and data workloads. You write Python, declare the image, GPU and resources in code, and Modal runs it as functions, web endpoints, scheduled jobs or isolated Sandboxes for code that agents generate. GPUs range from the T4 to the B300, and CPU and memory are metered separately in fractions of a core and GiB.

It competes with serverless GPU clouds such as RunPod Serverless and with model hosts such as Replicate, and more loosely with general serverless platforms. Our serverless platforms ranking covers the non-GPU side, and the cheapest GPU cloud guide compares hourly GPU rentals.

How Modal works

Billing starts when a container loads and runs while it processes inputs. By default Modal keeps a container alive for 60 seconds after the last input, and that time is billed; you can change the timeout. Once an app scales to zero, charges stop.

CPU is metered in physical cores (one core equals 2 vCPUs on AWS or GCP), with a minimum of 0.125 cores and 128 MiB of memory per container. You pay the higher of what you request and what you use, so over-requesting resources costs money even when the code sits idle. Starter and Team are pay-as-you-go with no prepaid credits required, and a workspace budget can cap monthly spend. Credits from AWS, GCP or Azure cannot be applied, though Enterprise customers can use committed spend through the AWS and GCP marketplaces.

Modal key features

Per-second billing with scale to zeroEssential
GPU, CPU and memory are each billed by the second, and nothing is charged once an app has no containers. For an endpoint that sees bursts of traffic, this removes the cost of a GPU sitting idle between requests, which is where hourly rentals waste money.
A GPU range from T4 to B300Essential
The rate card runs from the T4 at $0.000164 a second to the B300 at $0.001972 a second, with L4, A10, L40S, A100, H100, H200 and B200 in between. You can test a model on a cheap GPU and move up without changing platforms.
Starter and Team plans with monthly credit
Starter is $0 a month plus compute and includes a monthly credit, 3 seats, 100 containers and 10 concurrent GPUs. Team is $250 a month plus compute with $100 of credit, unlimited seats, 5,000 containers and 50 concurrent GPUs, plus custom domains, rollbacks and budgets.
Sandboxes for agent and untrusted code
Sandboxes run isolated environments for coding agents and RL rollouts, billed on active CPU and memory rather than reserved capacity. They use non-preemptible rates, 3 times the list CPU and memory price, so estimate from average usage, not peak.
Multipliers for region and non-preemptible runs
Pinning a region raises prices 1.15 to 1.75 times base, and non-preemptible execution costs 3 times. Teams with data residency needs or jobs that cannot be interrupted should price these in from the start.
Enterprise compliance and support
HIPAA, audit logs, SSO and private Slack support are Enterprise only, as is the 100 TiB of monthly egress. SOC 2 compliance is included on every plan, which covers most startup security reviews.

Modal pricing

Modal charges per second: L4 $0.000222, A10 $0.000306, L40S $0.000542, A100 80 GB $0.000694, H100 $0.001097 and B200 $0.001736, plus CPU at $0.0000131 per core per second and memory at $0.00000222 per GiB per second. Volumes cost $0.09 per GiB a month with 1 TiB free, and egress $0.04 per GiB beyond the plan's allowance. Prices were checked on November 4, 2026 on the Modal pricing page.

Scenario: an inference endpoint on one L40S, busy 4 hours a day for 30 days, with 2 physical cores and 16 GiB of memory requested (our workload assumption). That is 432,000 billed seconds. GPU: $234.14. CPU: $11.32. Memory: $15.34. Total $260.80, or $230.80 after Starter's monthly credit, before the 60-second keep-alive after each burst. All figures are our arithmetic on Modal's list rates.

On Replicate's L40S at $0.000975 a second, with 10 CPUs and 65 GB of RAM included, the same busy hours cost $421.20 if you run a public model. Deploy your own private model there and Replicate also bills setup and idle time for as long as the instance is online, so the gap widens. On RunPod Serverless, the 48 GB L40-class tier lists $1.75 an hour, $210.00 for 120 hours. Run the GPU around the clock instead and Modal's GPU charge alone is $1,404.86 a month, while a RunPod L40S pod at $1.09 an hour is $784.80. Per-second billing wins on bursts and loses on steady load.

Three gotchas move the bill: the keep-alive window is billed, CPU and memory are charged on the higher of request and use, and region selection or non-preemptible execution multiplies the rate. Rival prices come from the Replicate, RunPod and fal pricing pages, read the same day.

PlanPriceBest for
Starter$0/mo + compute$30 monthly credit, 3 seats, 100 containers, 10 concurrent GPUs, 1 TiB egress, 1-day logs.
Team$250/mo + compute$100 monthly credit, unlimited seats, 5,000 containers, 50 GPUs, rollbacks, budgets, custom domains.
EnterpriseCustomVolume discounts, higher GPU concurrency, HIPAA, SSO, audit logs, private Slack support.
Nvidia T4 / L4$0.000164 / $0.000222 per secEntry GPUs for small models and image work.
Nvidia L40S$0.000542 per sec48 GB GPU for mid-size LLM and diffusion inference.
Nvidia A100 80 GB / H100$0.000694 / $0.001097 per secLarge-model inference and fine-tuning.
Nvidia H200 / B200 / B300$0.001261 / $0.001736 / $0.001972 per secTop-end GPUs for the largest models.
CPU and memory$0.0000131 per core-sec; $0.00000222 per GiB-secMetered separately, minimum 0.125 cores and 128 MiB per container.

Modal pros and cons

What we like

  • Per-second billing and scale to zero, so idle time after the keep-alive costs nothing.
  • Monthly free compute credit on both Starter and Team.
  • GPUs from T4 to B300 on one rate card.

What could be better

  • At full utilization an L40S costs $1,404.86 a month in GPU time, more than an hourly pod.
  • Region selection adds 1.15 to 1.75 times and non-preemptible runs 3 times the list price.

Who Modal is for

Modal fits Python teams whose GPU usage comes in bursts: an inference API with uneven traffic, nightly batch jobs, fine-tuning runs and agent Sandboxes. The free monthly credit covers experiments, and per-second billing means a quiet week costs little. Startups can also apply for credit grants, and academics for up to $10,000 in credits.

Skip it for a model that serves traffic around the clock: an hourly pod on RunPod costs less at full utilization. Skip it too if you just want to call an existing open model: a hosted API is simpler, as our LLM APIs ranking shows, and Ollama runs models on your own machine for free. Non-Python stacks will find Modal's code-first model less natural than a container host such as Fly.io.

Best Modal alternatives

If Modal is not the right fit, these are the closest options.

ToolBest forStarts at
ModalPython teams running bursty GPU inference, batch jobs, fine-tuning or agent Sandboxes.Free plan · Team $250/mo · custom-priced plans tooVisit →
ReplicateDevelopers who want to call public models by API and pay only while a request runs.Per second by hardware: L40S $0.000975, H100 $0.001525, T4 $0.000225Visit →
RunPodTeams that want the lowest hourly GPU price, on pods or serverless endpoints.Serverless 48 GB tier $1.75/hrVisit →
falTeams building on generative image and video models who want hosted model APIs and their own deployments.Custom deployments from $4.00/hr (RTX PRO 6000) and $4.50/hr (H100) liVisit →
Replicate
Model hosting platform with thousands of public models and per-second hardware pricing.
Visit →
RunPod
GPU cloud with per-second pods, serverless endpoints and multi-node clusters in 30+ regions.
Visit →
fal
Generative media platform with per-output model APIs and hourly GPUs for custom apps.
Visit →

The bottom line

Modal is worth it for bursty GPU work written in Python, and the free credit makes it cheap to prove. At four busy hours a day, an L40S endpoint lands near the cost of RunPod Serverless and well below Replicate, with a smoother developer workflow. Once a GPU runs most of the day, rent a pod instead.

Compare GPU prices in our cheapest GPU cloud guide, serverless options in the serverless platforms ranking and AWS Lambda pricing, free allowances in cloud free tiers compared, and agent runtimes in the AI agent sandboxes ranking. Toolradar keeps profiles of Modal, Replicate and RunPod.

Methodology: Modal's rates, plan limits and billing rules were read on its pricing page on November 4, 2026, and rivals' rates on their pricing pages the same day. The workload is our assumption and the totals are our arithmetic on list rates before tax. Devshot does not benchmark platforms hands-on.

Cite this: Devshot, "Modal Review 2026: Pricing, Pros, Cons and Alternatives", November 2026.

Frequently asked questions

Is Modal worth it in 2026?

Yes for bursty GPU workloads in Python. Modal bills per second, scales to zero and includes monthly credit on Starter. For an L40S busy 4 hours a day it costs about $230.80 a month after credit by our arithmetic. For a GPU busy around the clock, an hourly pod elsewhere is cheaper.

How much does Modal cost?

Starter is $0 a month plus compute and Team $250 a month plus compute. Compute is per second: an L4 costs $0.000222, an L40S $0.000542 and an H100 $0.001097, plus CPU and memory. Prices were read on Modal's pricing page on November 4, 2026.

Does Modal have a free tier?

Starter includes $30 of free compute every month for up to 3 seats, with 10 concurrent GPUs and 1 TiB of egress. Usage above the credit is pay-as-you-go. Startups and academics can apply for larger credit grants.

What does Modal charge for idle time?

Nothing after an app scales to zero. Modal bills load time, processing time and a default 60-second keep-alive after the last input, which you can shorten. Once no containers are running, billing stops.

Modal vs Replicate: which is cheaper?

For your own model, usually Modal. Replicate's L40S costs $0.000975 a second against $0.000542 on Modal, and Replicate bills private models for setup and idle time while their instances are online. For calling public models by API, Replicate charges only while a request runs.
Related guides

Some offers on this page may be paid placements or contain affiliate links.

Get the Devshot brief

Free daily newsletter, read in 5 minutes.

Subscribe free