Replicate vs Baseten: hosted models, custom deployments and idle cost

Start by deciding whether you are calling a model someone else hosts or deploying your own. Both platforms support each, and the comparison changes with the mode.

If you need an existing image, video or audio model behind an API, Replicate's catalog is the broader starting point. Baseten's hosted Model APIs cover a curated set of language models priced per token. If you are deploying your own model on dedicated GPUs, both platforms charge for running capacity rather than for requests. The choice then turns on packaging (Cog or Truss), how much control you need over scaling and releases, and whether you need enterprise hosting options such as running inside your own cloud.

This comparison is based on each company's documentation and pricing pages, checked on September 24, 2026. It is not a benchmark, and it does not name a faster or cheaper platform for every workload.

Where the two platforms differ

Hosted model APIs cover different catalogs

Replicate's public models only charge for the time they spend processing your request. Setup and idle time are free, but you share a hardware pool with other customers, so you can hit cold boots or scaling limits when a model is busy. Its official models are kept warm and priced by output, such as images generated or tokens. Replicate billing, official models.

Baseten's Model APIs work differently. They expose OpenAI-compatible endpoints for a supported list of language models, billed per million input and output tokens. Default rate limits depend on your account tier: a verified Basic account starts at 120 requests and 500,000 tokens per minute, Pro raises the token limit to 1,000,000, and Enterprise limits are custom. Baseten Model APIs pricing and limits.

So a hosted-API comparison is only like for like when both catalogs carry your model. For an image or video generation feature, Baseten's route is a dedicated deployment, which changes the billing model entirely. For an LLM feature, compare the per-token price for your exact model, then check whether the rate limit covers your peak traffic.

A custom deployment bills for capacity, not requests

Replicate charges private models and deployments for all time an instance is online: setting up, idle and active. It lists an H100 at $5.49 an hour ($0.001525 a second) as of September 24, 2026. Fast-booting fine-tunes are the exception and bill only for active time. Replicate pricing.

Baseten bills each running replica by the minute and lists an H100 at $0.10833 a minute, about $6.50 an hour, on the same date. A deployment at zero replicas costs nothing, but starting a replica and loading the model is billable. Its pricing page says you don't pay for idle time, and then defines billable time to include deploying and scaling up or down. Baseten pricing, Baseten autoscaling.

The scale-down delay is where that matters. Suppose a hypothetical internal tool sends one minute of requests every hour to a Baseten H100 deployment that scales to zero. With the default 15-minute delay, each burst keeps the replica for at least 16 billed minutes, about $1.73, plus however long it takes to start. Across 24 bursts that is about $41.60 a day before startup time. This is arithmetic from list prices, not a measured bill. Shortening the delay cuts idle minutes and adds cold starts.

Don't compare the hourly rates in isolation either. The two pricing pages don't describe identical instances, and throughput per replica depends on the serving engine. Compare the bill for the same workload.

Cog versus Truss decides the migration work

Replicate builds custom models with Cog. You describe the environment in cog.yaml and write a Python class that runs the model; Cog generates a Docker image with an HTTP server. Because the output is a standard container, you can also run it on your own infrastructure. Deploy a custom model on Replicate.

Baseten uses Truss. For supported architectures, a config.yaml can declare the model, hardware and inference engine with no serving code. A Python model class handles custom preprocessing. A custom Docker server lets you bring an existing vLLM, SGLang or Triton container. Develop a model on Baseten.

Moving a custom model between them usually means rewriting the wrapper and retesting inputs and outputs. Baseten's custom Docker route may reduce that work for a container you already have, but check its requirements first: a standard build expects a Debian-based image with Python, reserves port 8080 and runs containers as a non-root user by default. You also declare readiness and liveness endpoints. Custom Docker containers. If you only call hosted models, you don't need Cog or Truss at all: switching means changing API calls and output handling.

Autoscaling controls and what they don't cover

Replicate deployments let you set minimum and maximum instances, switch hardware without code changes, and roll out new versions without downtime. Its documentation lists rolling updates and canary releases. Replicate deployments.

Baseten exposes more of the scaling decision. You set a concurrency target, target utilization, autoscaling window, scale-down delay and scale-down rate. A new deployment defaults to a minimum of zero and a maximum of one replica. When traffic exceeds the maximum, requests wait and can be rejected by load shedding. Environments separate staging from production, and rolling deployments shift traffic to a new version gradually. Baseten environments.

Two limitations deserve attention before you commit. Choosing a Baseten region does not reserve GPU capacity, and a deployment fails if the accelerator isn't available there. Regional deployments. A monthly budget can stop Model API requests, but it never stops dedicated deployments or training, which keep accruing charges. Baseten billing.

Enterprise features and company status

Baseten states SOC 2 Type II certification and HIPAA compliance on its pricing page and security documentation. Several features are Enterprise-only: self-hosted and hybrid deployments, custom SLAs, using existing cloud commitments, and custom global regions. Priority access to high-demand GPUs starts at Pro. Baseten security, hosting options.

Replicate's enterprise page mentions data processing agreements, SLAs, reserved compute and dedicated support, and its pricing page notes that additional multi-GPU capacity is available through committed spend contracts. The pages checked did not name a SOC 2 report or HIPAA terms, so ask sales for documentation if you need them. Replicate enterprise.

Cloudflare acquired Replicate on December 1, 2025, according to Cloudflare's quarterly filing. Replicate's documentation and pricing still run on replicate.com as of September 24, 2026. Check the current terms if long-term support matters to your decision. Cloudflare Form 10-Q.

Test both with the same workload before you standardize

Run one representative workload on each platform with a spending limit you control:

• Use the same model, inputs and output settings.
• Send a request after a quiet period and record the time until your app has a usable result.
• Send a burst that reaches your maximum replica or instance count, and note which requests wait or fail.
• Compare billed seconds or minutes in the usage dashboard with the time spent processing requests.
• Record how long packaging and deployment took, including the changes your app needed.

If Replicate hosts the model you need, start there and check what idle capacity would cost once you move to a deployment. If you are serving your own model and need detailed scaling controls or a path into your own cloud, Baseten deserves the closer evaluation. Let the test bill and response times decide close cases.