Modal Pricing 2026: Per-Second GPU Rates, Starter vs Team and Real Inference Cost

Short answer: Modal bills compute by the second with no minimum. An Nvidia H100 costs $0.001097 a second, about $3.95 an hour, an L4 about $0.80 and a B200 about $6.25. The Starter plan is free with $30 of compute a month; Team costs $250 a month plus compute, with $100 of credits. Checked on Modal's pricing page, 30 October 2026.

Per-hour, Modal is not the cheapest GPU you can rent. It wins when your GPU would otherwise sit idle: you pay only while a container loads or works, plus a short keep-alive window. The utilisation scenarios below show where that stops being true.

Modal pricing at a glance

From Modal's pricing page, read on 30 October 2026. Compute is billed on top of the platform fee once included credits run out.

The Team fee buys operations features more than compute: custom domains, a static IP proxy, deployment rollbacks (3 versions), environment-level budgets, RBAC, 30-day log retention instead of 1 day, unlimited scheduled functions instead of 5 crons, and 10 TiB of free egress. After its $100 credit, Team costs $150 a month more than Starter for a team that uses at least that much compute (our arithmetic).

How much does Modal cost per GPU hour?

Modal publishes per-second prices. The hourly figures are those rates × 3,600:

Per-second rates from Modal's pricing page, 30 October 2026.

A GPU container also bills CPU and memory. A physical core (2 vCPUs) costs $0.0000131 a second and memory costs $0.00000222 per GiB a second, charged on whichever is higher, what you request or what you use. A container with 2 cores and 16 GiB adds about $0.22 an hour to any GPU (our arithmetic), so budget about $4.17 an hour for a typical H100 endpoint rather than the headline rate.

Location and reliability cost extra. Choosing a specific region multiplies prices by 1.15 to 1.75, and non-preemptible execution, which guarantees a function will not be interrupted, applies a 3x multiplier to the CPU and memory list prices. Sandboxes always bill CPU and memory at that 3x rate.

What does a Modal model endpoint cost per month?

Here is one H100 endpoint at different levels of use over a 30-day month (720 hours), against a reserved H100 that bills around the clock:

Our arithmetic from Modal's per-second rates, before included credits, region multipliers or egress. Billable hours include load time and the keep-alive window.

For comparison, Together AI rents a reserved H100 in a GPU cluster at $3.45 an hour for 31 to 90 days or $3.19 for 91 to 180 days, which is $2,484.00 or $2,296.80 for the same 720 hours (our arithmetic from Together AI's pricing page, same date). Our Together AI pricing breakdown covers its serverless and dedicated options too.

Dupple data: on GPU cost alone, a Modal H100 stays cheaper than a 91-to-180-day reserved H100 at Together AI until it is busy about 81% of the time, and cheaper than a 31-to-90-day reservation until about 87%, our arithmetic from both vendors' pricing pages on 30 October 2026.

The practical reading: an endpoint that serves office hours, a demo, or traffic that spikes and falls costs a fraction of a reserved GPU on Modal. One that stays busy almost all day is cheaper reserved, provided you can run the serving stack yourself. Modal's own example on the pricing page makes the same point with 75 reserved GPUs against an average of 50 busy ones.

What counts as billable time on Modal?

Modal bills from the moment your application starts loading until the container stops processing, plus a keep-alive window after the last input. That window defaults to 60 seconds and is configurable. Once an app scales down to zero containers, charges stop.

The keep-alive is the setting most worth tuning. A shorter window cuts idle billing for sparse traffic but means more cold starts, each of which bills its load time again. Large models that take a long time to load can make frequent cold starts more expensive than a longer keep-alive.

Modal also offers Shared Endpoints, billed per token rather than per second. Since 1 September 2026, token usage is billed from the first request and is not covered by the Starter plan's $30 of included compute, so a free-tier experiment with a shared endpoint can still produce an invoice.

What do Modal's included credits actually buy?

The $30 Starter credit covers about 7.6 hours of H100 time or about 50 hours on a T4, GPU only (our arithmetic). The $100 Team credit covers about 25 H100 hours. These are monthly allowances for prototyping, so treat them as a discount on a production bill rather than a budget for one.

Modal's own examples give a sense of small-workload costs. A daily one-hour job using half a CPU core and 256 MiB of memory costs $0.77 over 30 days. Stable Diffusion on an A10G, at about 1.5 seconds an image, costs about $0.491 per 1,000 images. Those figures fit easily inside the free credit.

Storage and network are cheap until a project moves large model weights or datasets around. Volumes cost $0.09 per GiB a month with 1 TiB included, and network egress costs $0.04 per GiB beyond the plan's free allowance (1 TiB on Starter, 10 TiB on Team, 100 TiB on Enterprise).

The costs that surprise Modal users

• CPU and memory ride along. Every GPU container also bills cores and memory, at the higher of requested or used.
• Keep-alive bills. The default 60 seconds after the last request is charged on every scale-down.
• Region pinning costs 15% to 75% more. Leave region selection off unless data residency requires it.
• Non-preemptible means triple CPU and memory. Sandboxes always pay that rate.
• Shared Endpoint tokens are billed from the first request, even on Starter.
• Seat cap on Starter. More than 3 people in a workspace requires Team.

Modal's workspace budgets and spend limits cap monthly charges, and setting one is worth the minute it takes before you deploy anything public.

Are there discounts or free credits?

Yes. Modal's pricing page lists credit grants for early-stage startups, up to $10,000 in compute credits for graduate students, labs and researchers, and purchasing through the AWS and GCP marketplaces to use committed cloud spend. Credits from AWS, GCP or Azure cannot be applied directly, though AWS Activate startups can qualify for extra Modal credits through the startup program. Enterprise adds volume-based discounts.

How we checked these prices

We read Modal's pricing page, including its plan comparison and FAQ, and Together AI's pricing page for the reserved-GPU comparison, on 30 October 2026. Hourly rates, monthly endpoint costs, break-even utilisation and credit coverage are our own arithmetic from those per-second and per-hour list prices, using a 720-hour month.

For other ways to run models, see the best free serverless platforms, MLOps tools and AI cloud cost optimization tools. If you would rather not run GPUs at all, LLM gateways route requests to hosted models, and free local LLM tools run small ones on your own machine. Training a custom model first? What it costs to fine-tune an LLM covers that side of the budget.

FAQ

Bottom line

Use Modal for endpoints and jobs that are idle much of the time: at 10% to 40% utilisation, per-second billing costs a fraction of a reserved GPU. Tune the keep-alive window, set a budget, and price in CPU and memory. Once a GPU is busy around four-fifths of the day, a reserved instance becomes the cheaper home.

Cite this: Dupple, "Modal Pricing 2026: Per-Second GPU Rates, Starter vs Team and Real Inference Cost", November 2026.

Pricing out the AI stack for your team? Start the Dupple X trial.