Modal
FreemiumAI infrastructure that developers love.
Published 1 October 2026
Scores
Popularity4/5
A $4.7B valuation, a few hundred million dollars in annualised revenue, and customers such as Anthropic and Meta make it one of the best-known AI infrastructure names among ML engineers and agent builders, though web-focused developers rarely touch it.
Learning Curve2/5
A decorator and a few lines of Python deploy a function or start a sandbox with no Docker or cloud console, although tuning images, GPUs, and concurrency for production takes more understanding of the platform.
Flexibility5/5
Arbitrary container images, any GPU class, sandboxes, web endpoints, schedules, and storage primitives let one platform cover agent code execution, inference, training, and data jobs.
Performance5/5
Sub-second container starts, fast GPU cold starts, and autoscaling that has handled more than a million sandboxes in a weekend put it at the top of the serverless compute field.
Portability2/5
Code is plain Python and images are standard containers, but the decorators, sandbox API, and scheduling model are Modal-specific and there is no way to run the platform elsewhere.
About Modal
Modal is a serverless compute platform for AI and data workloads. Developers decorate Python functions with the hardware, container image, and dependencies they need, and Modal runs them in the cloud on CPUs or GPUs, scaling from zero to thousands of containers in seconds and billing per second. The same platform serves model inference endpoints, batch jobs, fine-tuning, scheduled tasks, and notebooks, with volumes and buckets for storage.
For agents, the relevant product is Modal Sandboxes: containers created at runtime through the Python or JavaScript SDK to run code a model generated. Each sandbox is isolated with gVisor, a user-space kernel that intercepts system calls, starts in under a second from cached images, and lives for five minutes by default or up to 24 hours, with idle timeouts, filesystem snapshots, tunnels for exposing ports, and optional GPUs. Sandboxes have become one of Modal's biggest businesses: the platform has launched more than a billion of them, and Lovable used Modal to run over a million sandboxes during a single launch weekend. It is one of the built-in sandbox providers in the OpenAI Agents SDK.
The main reason to choose Modal over a sandbox-only provider is GPU access and having one platform for everything around the agent: the sandbox that runs generated code, the GPU endpoint that serves a fine-tuned model, and the batch jobs that process data can all share code, images, and volumes.
Modal is a closed-source, cloud-only service from Modal Labs, which raised a $355M Series C in 2026 and counts Anthropic, Meta, Cognition, and Ramp among its customers. Pricing is usage-based with free monthly credits on the Starter plan. Teams that need self-hosting or microVM-level isolation usually pick E2B instead, and Daytona leads on cold-start speed.
Key Features
- Sandboxes for running AI-generated code, isolated with gVisor
- Serverless CPU and GPU functions defined in Python code
- Scale from zero to thousands of containers in seconds
- Sandbox lifetimes up to 24 hours with idle timeouts and snapshots
- GPU-backed sandboxes and inference endpoints on the same platform
- Volumes, buckets, cron jobs, and notebooks for surrounding workloads
- Python and JavaScript SDKs with per-second billing
Pros
- One platform covers sandboxes, GPU inference, and batch jobs with shared code and images
- Proven at very large sandbox volumes, including million-sandbox launch spikes
- Infrastructure is defined in Python, with no YAML, Dockerfiles, or Kubernetes required
- Free monthly credits cover small projects and experiments
Cons
- Cloud-only and closed-source, with no self-hosted option
- gVisor containers share a host kernel, a weaker boundary than E2B's Firecracker microVMs
- Python-first design; the JavaScript SDK covers less of the platform
- Per-second GPU billing can get expensive for always-on workloads compared with reserved servers
Modal Pricing
Freemium- · $30 of free compute each month
- · Up to 3 seats, 100 containers, and 10 concurrent GPUs
- · $100 of compute credits included
- · Unlimited seats, 1,000 containers, and 50 concurrent GPUs
- · Custom limits, volume discounts, and support
- · Quoted by sales
Tools Related to Modal
Works well with Modal(1)
Modal's serverless GPUs are a common place to run vLLM without owning hardware: a vLLM server is defined in Python and scales to zero between requests.
Integrates with Modal(2)
Modal is a built-in sandbox provider in the OpenAI Agents SDK: sandbox agents run their commands and file edits inside Modal Sandboxes through the SDK's adapter.
Hermes Agent includes Modal as one of its terminal backends, so the agent's shell commands run in serverless Modal containers instead of on the host machine.
Alternatives to Modal(2)
E2B is a sandbox-only platform with microVM isolation and a self-hosting option; Modal runs sandboxes in gVisor containers on a broader serverless platform with GPUs, inference endpoints, and batch jobs.
Daytona specialises in agent sandboxes, with the fastest starts and environments that can persist indefinitely; Modal offers sandboxes as one part of a serverless compute platform, with GPUs and model serving alongside.
Vendor
Modal Labs
Website →Tags
Details
- Maintained
- Yes