Baseten vs Modal: where should you run your own models in 2026?
Independent research, updated 2026-11-09. Figures come from Baseten's and Modal's own pricing pages. Approximate hourly rates for Modal are our arithmetic from its per-second prices. We did not run hands-on tests.
Answer in brief: Baseten is a managed inference platform. It offers Model APIs with the lowest gpt-oss-120b price we found ($0.10 input / $0.50 output per 1M tokens), Fast variants of popular models, and Dedicated Inference on GPUs billed by the minute (H100 at $6.50 per hour). Modal is serverless compute for your own Python code. Its GPU list prices are lower (H100 at about $3.95 per hour, billed per second, with scale to zero), but you run the serving stack yourself. Region selection costs 1.15–1.75x and non-preemptible execution costs 3x.
Top pick for production model serving: Baseten. You get token APIs and dedicated GPUs on one bill, plus volume discounts, reserved and priority compute, custom regions and SLAs at production scale.
Best for code-first teams with spiky traffic: Modal. It bills per second, scales to zero and includes $30 a month in free credits on Starter.
The rate card, side by side
| Item | Baseten | Modal |
|---|---|---|
| What it is | Model APIs, Dedicated Inference, Training, Sandboxes | Serverless GPUs and CPUs for your code, plus Sandboxes, Training and featured model endpoints |
| H100 | $0.10833/min (about $6.50/hr) | $0.001097/sec (about $3.95/hr) |
| B200 | $0.16633/min (about $9.98/hr) | $0.001736/sec (about $6.25/hr) |
| A100 80GB | $4.00/hr | about $2.50/hr |
| L4 / T4 | $0.85 / $0.63 per hour | about $0.80 / $0.59 per hour |
| Token APIs | gpt-oss-120b $0.10 / $0.50; GLM-5.3 $1.40 / $0.14 cached / $4.40; Kimi K3 $3.00 / $15.00 | Featured endpoints (Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017, GLM 5.3, Muse Spark Open); not priced on the pricing page |
| Speed variants | Fast versions, for example GLM-5.3 Fast at $2.10 / $6.60 | Your own engine and configuration |
| Premiums | Custom regions on production contracts | Region selection 1.15–1.75x; non-preemptible 3x |
| Plans | Pay as you go; production-scale custom terms | Starter $0 with $30/mo credits; Team $250/mo with $100/mo credits; Enterprise custom |
| Concurrency caps | Set by capacity agreements at scale | 10 GPUs on Starter, 50 on Team |
| Compliance | Publishes a Service Level Agreement and a security page | SOC 2 on all plans; HIPAA, SSO and audit logs on Enterprise |
| Procurement | Volume discounts | AWS and GCP marketplace committed spend (Enterprise) |
Why Baseten wins for shipping models
- Two products in one account. Prototype on Model APIs, then move hot or custom models to dedicated GPUs without changing vendors.
- The lowest gpt-oss-120b list price we saw. $0.10 / $0.50 per 1M tokens, compared with $0.15 / $0.60 at most GPU clouds.
- Production terms on request. Volume discounts, reserved and priority compute, capacity planning, custom regions and SLAs are available through sales.
Why Modal wins for builders
- The cheapest GPU-seconds of the pair. At list price, an H100 on Modal is about 39% cheaper per hour than on Baseten.
- No idle bill. You pay only for compute time, by the CPU cycle and GPU-second, and Modal says it is more cost-effective than reserved compute for spiky workloads.
- Free runway. Starter includes $30 a month in credits. Startups can apply for credits, and academics can get up to $10k.
When to pick which
Pick Baseten if you want managed serving with per-token APIs beside your dedicated deployments, or if you need negotiated capacity and SLAs. Pick Modal if your team writes its own serving code (vLLM, SGLang, diffusion pipelines), your traffic is bursty, and you can live with preemptible defaults or will pay the multipliers for a fixed region or non-preemptible execution.
How each one bills you
Baseten: per 1M input, cached-input and output tokens on Model APIs. Dedicated Inference and Training are billed per minute of GPU or CPU time, and Sandboxes per vCPU-hour ($0.0655) and GB-hour of RAM ($0.0211). Modal: per second of GPU, CPU core and memory, plus an optional platform plan. Volumes cost $0.09 per GiB per month beyond 1 TiB free, and egress costs $0.04 per GiB beyond your plan's allowance.
FAQ
Is Modal cheaper than Baseten?
Per GPU-hour at list price, yes: an H100 costs about $3.95 on Modal versus $6.50 on Baseten. Modal's region and non-preemptible multipliers can close that gap, and Baseten's Model APIs avoid GPU billing altogether.
Does Modal offer per-token model APIs?
It lists featured model endpoints, but its pricing page doesn't publish per-token rates for them. Baseten publishes per-token Model API prices.
Which has better compliance options?
Modal publishes SOC 2 on all plans and HIPAA on Enterprise. Baseten publishes an SLA and a security page. Check each provider's trust documentation for your requirements.
Can I train on either?
Yes. Both sell training compute. Baseten's training rates match its dedicated GPU rates, and Modal bills training per second like its other compute.
Keep exploring
Back to the ranked list · Baseten review · Modal review · How to choose · Pricing comparison
·
Sources
- Baseten pricing
- Baseten Service Level Agreement
- Modal pricing
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.