Serverless GPU comparison - checked October 21, 2026

Best Serverless GPU in 2026: compare the billing contract first

RunPod worker seconds, Together AI tokens and Lambda GPU hours are not the same unit. This guide separates the deployment models before comparing the providers.

Choose by workload, not by a misleading starting price

Variable custom containers

Start with RunPod Flex when scale to zero matters. Model loading and idle timeout remain part of the bill.

Supported open-model API

Compare Together AI, CoreWeave and Nebius by the exact model, input tokens, cached tokens and output tokens.

Distributed container fleet

Evaluate Salad by running replicas, hardware priority and queue design, not by a generic serverless starting price.

Training or steady utilization

A dedicated Lambda instance can be the clearer contract when a GPU must stay allocated for a long job.

Serverless GPU decision matrix

The table intentionally has no cheapest-to-most-expensive ranking. Each row uses the unit published for that product. Source pages were checked on October 21, 2026.

Provider and contract	Billing unit	Idle behavior	Best fit	Important tradeoff	Official source	Provider
R RunPod Serverless Custom container workers	Per second while a worker runs, rounded to the next second. Storage is separate.	Flex workers scale to zero. Active workers stay online.	Custom model handlers with variable request volume	Startup, execution and idle-time phases are all billable.	RunPod Serverless pricing	Check provider
T Together AI Hosted model API	Per model and per unit, including separate input, cached-input and output token rates for chat models.	No replicas to size for the shared serverless catalog.	Calling supported open models without managing containers	The shared catalog, model availability and rates can change.	Together AI serverless model catalog	Check provider
S Salad Container Engine Distributed container groups	Per second while container instances run, at a rate set by hardware and priority.	Do not assume automatic scale to zero. Replica and queue behavior must be designed for the workload.	Cost-sensitive stateless container jobs that tolerate distributed capacity	Operational behavior differs from a managed per-request model API.	Salad Container Engine billing	Check provider
C CoreWeave Serverless Inference Hosted model API	Per input and output token for the curated serverless model catalog.	No cluster to provision for the serverless catalog.	Agents, RAG and model evaluation on supported models	Serverless accepts catalog models and LoRA weights. Full custom-weight deployments use a dedicated GPU-time contract.	CoreWeave Serverless Inference	Check provider
N Nebius Token Factory Hosted model API	Per input and output token, with model-specific rates.	No idle GPU charge for the shared serverless API.	Open-model APIs where a shared global endpoint is acceptable	Public endpoints can change processing region without notice. Use a dedicated endpoint when region stability matters.	Nebius Token Factory pricing	Check provider
λ Lambda GPU Cloud Dedicated on-demand GPU instance	Per GPU-hour with minute-level billing for the instance.	No. The instance remains a provisioned GPU until you stop it.	Training, long jobs or steady inference that needs full instance control	This is the deliberate non-serverless fallback, not a scale-to-zero product.	Lambda GPU instances	Check provider

How to compare the total cost

Container worker

Estimate starts, model-load seconds, execution seconds, post-request idle seconds, concurrent workers and persistent storage. RunPod explicitly bills the startup, execution and idle-time phases.

Hosted model API

Estimate requests multiplied by input, cached-input and output tokens. Use the exact model row. Do not convert a token API to a GPU-hour headline without a measured throughput assumption.

Dedicated GPU

Estimate allocated minutes or hours, attached storage and any data charges. Dedicated capacity can be simpler for training and steady traffic even when it does not scale to zero.

Method boundary: This is a comparison of current published service contracts, not a latency benchmark. We do not claim a universal winner or reproduce promotional prices that can change independently.

Six options, with the product boundary made explicit

RunPod Serverless

Custom container workers

Use it for: Custom model handlers with variable request volume
Billing: Per second while a worker runs, rounded to the next second. Storage is separate.
Watch for: Startup, execution and idle-time phases are all billable.

Read official terms Visit RunPod Serverless

Together AI

Hosted model API

Use it for: Calling supported open models without managing containers
Billing: Per model and per unit, including separate input, cached-input and output token rates for chat models.
Watch for: The shared catalog, model availability and rates can change.

Read official terms Visit Together AI

Salad Container Engine

Distributed container groups

Use it for: Cost-sensitive stateless container jobs that tolerate distributed capacity
Billing: Per second while container instances run, at a rate set by hardware and priority.
Watch for: Operational behavior differs from a managed per-request model API.

Read official terms Visit Salad Container Engine

CoreWeave Serverless Inference

Hosted model API

Use it for: Agents, RAG and model evaluation on supported models
Billing: Per input and output token for the curated serverless model catalog.
Watch for: Serverless accepts catalog models and LoRA weights. Full custom-weight deployments use a dedicated GPU-time contract.

Read official terms Visit CoreWeave Serverless Inference

Nebius Token Factory

Hosted model API

Use it for: Open-model APIs where a shared global endpoint is acceptable
Billing: Per input and output token, with model-specific rates.
Watch for: Public endpoints can change processing region without notice. Use a dedicated endpoint when region stability matters.

Read official terms Visit Nebius Token Factory

Lambda GPU Cloud

Dedicated on-demand GPU instance

Use it for: Training, long jobs or steady inference that needs full instance control
Billing: Per GPU-hour with minute-level billing for the instance.
Watch for: This is the deliberate non-serverless fallback, not a scale-to-zero product.

Read official terms Visit Lambda GPU Cloud

One notable alternative not forced into the affiliate table

Modal is a container-oriented serverless platform with per-second CPU, memory and GPU metering. Its current pricing page also separates workspace plans from compute usage. We include it as a research alternative, not as a ranked row, because this page does not currently have a verified Filedesign destination for Modal. See Modal's official pricing.

Frequently asked questions

What does serverless GPU mean?

The label covers different contracts. A platform may run your custom container and scale workers, or expose a hosted model API billed by tokens. A dedicated GPU instance is a separate model because the customer keeps the GPU allocated.

Which serverless GPU option scales to zero?

RunPod documents that Flex workers scale to zero while Active workers stay online. Hosted model APIs hide replica management from the customer. Do not assume every container platform reaches zero running replicas without checking its scaling configuration.

Which serverless GPU provider is cheapest?

There is no honest single winner without a workload. Token APIs, running container seconds and dedicated GPU hours measure different things. Estimate request volume, input and output tokens, execution time, startup time, idle timeout and storage before comparing totals.

Can I deploy my own model weights?

RunPod supports custom container workers. Hosted model APIs expose a catalog and may support adapters or dedicated endpoints under a different contract. Verify the exact model and deployment tier before choosing a provider.

Should I use serverless GPU for training?

Usually not for long training runs. Training commonly benefits from a dedicated GPU that stays allocated, predictable storage and checkpoint control. Serverless is strongest for inference or short jobs with variable demand.

Primary sources

Last source check: October 21, 2026. Recheck the exact model catalog, billing unit and account terms before spending.