Modal review

Modal is cloud infrastructure for teams building AI applications. According to its homepage, it provides on-demand CPU and GPU compute for inference, model training, sandboxes, and other compute-heavy workloads, with usage-based billing.

Its feature set covers several stages of an AI workload: hosting inference apps, running fine-tuning and reinforcement learning jobs, and creating sandboxes for coding agents or reinforcement learning rollouts. It also runs general-purpose code, asynchronous batch jobs, and HTTPS endpoints that can idle at zero. Compute can scale up when needed and back down when work finishes.

Modal includes logs, metrics, audit logs, role-based access controls, and network and data residency controls. The pricing page lists a free Starter plan and a $250-per-month Team plan. That monthly fee is not the whole bill: compute is charged separately by usage.

Modal brings inference, training, agent sandboxes, and general-purpose compute under one service. That breadth could suit teams that would otherwise stitch together separate infrastructure for model serving, batch jobs, and development environments. The product supports custom model-serving stacks, while training workloads range from single-GPU to multi-node, according to the published product information.

The scale-to-zero model is a useful cost control for workloads that do not need to stay online. Modal says serverless functions and HTTPS endpoints can idle at zero, and CPU and GPU capacity can scale back down when work finishes. It also offers operational controls beyond compute: the feature list includes logs, metrics, granular RBAC, audit logs, and network and data residency controls.

The free Starter plan gives small teams a way to begin without a monthly plan fee. The pricing page lists $30 in monthly compute credits, up to three seats, 100 containers, GPU concurrency of 10, 1 TiB of monthly free network egress, and five deployed crons. Those allowances make the free tier concrete, though they are not a substitute for checking whether a workload fits its caps.

The main pricing catch is that the $250-per-month Team fee does not buy unlimited compute. The pricing page lists $100 in monthly compute credits, then bills compute by usage. Team includes unlimited seats, 5,000 containers, GPU concurrency of 50, 10 TiB of free monthly network egress, and unlimited Scheduled Functions. Starter is smaller: three seats, 100 containers, GPU concurrency of 10, 1 TiB of egress, and five deployed crons. Enterprise pricing is not published.

Metered rates add uncertainty: CPU is $0.0000131 per core-second, memory is $0.00000222 per GiB-second, and an H100 is $0.001097 per second; other GPU rates vary by model. Egress beyond the plan allowance costs $0.04/GiB, and volumes cost $0.09/GiB per month after the included 1 TiB. Region selection costs 1.15–1.75× base prices; non-preemptible execution costs 3×. Shared Endpoint token rates are not listed, though billing starts with the first request. The pricing page lists monthly fees, not annual pricing.

Modal has a free Starter plan, but its listed paid tier starts at $250/month, before usage charges. Starter costs $0/month and includes $30 in monthly compute credits, up to 3 seats, 100 containers, 10 GPU concurrency, 1 TiB of monthly free network egress, and 5 deployed crons. Team costs $250/month and includes $100 in compute credits, unlimited seats, 5,000 containers, 50 GPU concurrency, 10 TiB of free egress, and unlimited Scheduled Functions. That is a sharp jump if you only need more seats or capacity; the published pricing lists no intermediate paid tier. Enterprise pricing is custom.

The base fee is not an all-in compute bill. CPU costs $0.0000131/core/sec, memory $0.00000222/GiB/sec, and an H100 $0.001097/sec. Egress beyond plan allowances is $0.04/GiB; volumes after the included 1 TiB cost $0.09/GiB/month. Region selection adds 1.15–1.75× base prices, and non-preemptible execution costs 3×. Monthly plan fees are listed; annual pricing isn't published.

Switching to Modal is most straightforward when you can move one workload at a time: its listed use cases cover inference, training, sandboxes, serverless functions, and batch jobs. Budget for code changes wherever your app depends on another provider’s deployment, job, or autoscaling model. Modal’s pricing includes container, GPU-concurrency, and scheduled-function limits that may not map neatly to your current setup. Treat portability as a design choice, not a promise: the published material does not document a migration path, data-export tooling, or compatibility guarantees. Confirm how you will retrieve datasets, logs, model artifacts, and secrets before moving production workloads.

We can’t responsibly name common reasons customers leave based on Reddit: no Reddit posts or comments were included in the published material. The pricing does point to a possible exit trigger for cost-sensitive teams: a $250/month Team fee plus metered compute, with region and non-preemptible multipliers. If predictable monthly spend matters, model those charges before committing.