Replicate

Replicate lets you run models through an API without managing servers. Most public models bill by the second on the hardware they use, and some bill per image, per second of video or per token. On 31 October 2026 an H100 was $5.49 an hour and an A100 80 GB $5.04. It is quick to start and costs more per GPU hour than a pod, in exchange for zero setup.

Some links on this page go through /go/ and may earn us a fee; it never changes our order.

Facts at a glance

Compare with other hosts in the interactive hosting compare, and see the AI hosting overview for how it fits among GPU clouds and per-token APIs.

How Replicate works

You pick a model from the catalogue, send inputs through the API or a client library, and get outputs back. The platform runs the model on a GPU, bills you for the compute and shuts the hardware down when nothing is running. The catalogue includes public models contributed by Replicate and the community. You can also package your own model, which Replicate's docs describe with Cog, an open-source packaging tool, and deploy it either as a private model or as a deployment you scale yourself.

Two kinds of model behave differently on your bill.

Public models. Shared infrastructure. The pricing page says you pay only for the time it takes to process your request, whether you are running or training a public model. Cold boot time is the platform's problem, not yours, though it can add latency.

Private models and deployments. Dedicated hardware, so you do not share a queue with anyone. That means you pay for all the time instances are online: setup, idle waiting and active processing. The exception on the pricing page is fast-booting fine-tunes, which bill only for active time.

The prices

Hardware rates from replicate.com/pricing, checked on 31 October 2026. Hourly figures are the page's own.

Model-level prices on the same page include examples such as $0.04 and $0.025 per output image for two image models, $3.00 per thousand output images for another, and $0.09 or $0.25 per second of video output. Token-billed language models list rates per million input and output tokens, for instance $3.75 per million input tokens on one model. Rates differ by model and change often, so read the model's own page.

What you pay in practice

The volumes are examples we chose.

Example one: 10,000 images a month at $0.04 each. That is $400. At $0.025 each it is $250. Compare it with running an image model yourself on a Runpod Community RTX 4090 at $0.34 an hour: if a generation takes 4 seconds, 10,000 images is about 11 GPU hours, or about $3.78 of compute, before setup, queueing and idle time. Self-hosting can be many times cheaper at volume, and Replicate is cheaper at low volume where the idle hours would dominate.

Example two: a private model on an A100 80 GB, always on. 720 hours × $5.04 = $3,628.80. The same card on a Runpod Secure pod is $1.79 an hour, about $1,289 for the month. The premium is what you pay for managed scaling and the API.

Example three: a bursty app with 40 hours of real GPU work a month on an H100. 40 × $5.49 = $219.60, billed only for the processing time if the model is public and billed by time. On a rented H100 you would pay for every idle hour, so Replicate wins at this duty cycle.

Example four: a fine-tune. A training job of 3 hours on an H100 is 3 × $5.49 = $16.47, on top of any setup.

The break-even against a pod comes when the GPU is busy a large share of the month. Take the A100 80 GB: $5.04 per hour on Replicate against $1.79 on a Runpod Secure pod. A pod costs $1,289 for a month. Replicate reaches that at about 256 hours of busy time, which is roughly 35 percent of the month. Below that, per-second billing is cheaper. Above it, a pod is cheaper, and you also get the idle hours.

Deployments and scaling

For your own models, a deployment adds control. The docs describe auto-scaling from zero to hundreds of instances based on traffic, always-on instances to remove cold starts, scale-to-zero to cut cost, and instance health views. A deployment lets you set a maximum number of instances to cap spend and a minimum to keep some warm. That cap is the single most useful setting for protecting your budget from a traffic spike or a runaway loop.

Data and privacy

The data retention page says that for predictions created through the API, all input parameters, output values, output files and logs are automatically removed after an hour by default, and that you must save your own copies before then. For predictions made on the website, retention differs and you can delete them manually from the dashboard.

That is a short retention window. It is not a zero-retention guarantee, and it does not describe what happens on the underlying infrastructure during processing. For regulated data, ask for the data processing terms.

One more point: public models are written by third parties. Their code runs on Replicate's hardware, and the model's behaviour and licence are the author's. Read the licence on the model page before commercial use.

Limits

Replicate, Modal and Runpod Serverless side by side

All three bill by the second without a server to manage, but they sit at different points between convenience and control.

The H100 rate difference is large. Replicate's $5.49 is about 39 percent above Modal's $3.95 and about 104 percent above Runpod Community's $2.69 for a pod. That premium buys the catalogue and the lack of setup. If your model is one you will run for months, the premium is not worth paying. If you are testing ten models this week, it is. Read Modal and Runpod for the other two.

Cold boots and latency

Per-second billing implies that hardware is not running when nobody is calling. The first request after quiet therefore waits for hardware and model load. The pricing page notes that public models bill for processing time only, so boot time is not on your bill, but the delay lands on your user. For interactive features, that matters more than the cost. The deployment docs describe always-on instances to remove cold starts, which you pay for around the clock. The trade is the usual one: lower latency, higher cost.

If your product shows a spinner anyway, such as an image generator where people expect to wait, cold boots are acceptable. If it is a chat reply, they are not, and a token API or an always-warm instance is the better choice.

Using Replicate from an agent

An agent that calls a model through Replicate needs the same care as any outside tool.

• Store results immediately. Output files are removed after an hour by default for API predictions, so the agent should copy them to your own storage in the same step.
• Handle async. Many models take seconds or minutes. Use the polling or webhook route, and set timeouts so a stuck prediction does not hold a loop open.
• Cap spend. Set a maximum instance count on deployments, and give the agent a budget per task. A loop that calls an expensive video model 500 times is a real bill. At $0.25 per second of video output, 500 five-second clips would cost $625.
• Keep the token safe. One API token per environment, never in a prompt or a repository.
• Log the model version. Outputs differ between versions, and you will want to reproduce them.

For the wider pattern of agents and tools, see how to build an AI agent.

Cost controls that work

1. Prefer per-output pricing when you can predict it. A model billed at $0.04 per image gives a fixed cost per feature use.
2. Set max instances on any deployment. It is the cap on a runaway bill.
3. Watch the balance. The docs say stronger rate limits apply as credit runs low, to protect you from arrears, so top up before a launch.
4. Compare monthly against a pod every quarter. The break-even for an A100 80 GB is about 256 busy hours a month.
5. Review public model prices. They are set per model and change.

The Cloudflare news

Replicate's blog features a post titled "Replicate is joining Cloudflare". We have not read the terms of that arrangement, and the pricing page we read still shows Replicate's own rates. If you are building on Replicate for the long term, check the current announcement for what changes in pricing, regions and API stability. . Cloudflare's own inference product is covered on Cloudflare Workers AI.

Who should pick Replicate

Alternatives

• Modal: per-second billing for code you write, H100 at about $3.95 an hour.
• Runpod: pods and serverless, lower hourly GPU rates.
• Lambda Cloud: plain GPU instances, no egress fees.
• Hugging Face Inference: dedicated endpoints and a provider router.
• AWS Bedrock: language models inside AWS.
• Gateways: see OpenRouter alternatives.

Getting started

1. Create an account and add a payment method, since granted credit alone is rate limited.
2. Pick a model and read its page for price, licence and input fields.
3. Call it from the client library or a plain HTTP request, with your API token in an environment variable.
4. Save outputs right away, since API data is removed after an hour by default.
5. If traffic grows, create a deployment, set a maximum instance count, and compare the monthly cost with a pod.

Our take

Replicate trades GPU price for convenience. At $5.49 an hour for an H100 it is dearer than a pod, and at low and bursty use that does not matter, because you pay only while models run. We would use it for prototypes, for image and video features and for apps whose traffic we cannot predict. Once a model runs more than a third of the month, price a pod. Watch the Cloudflare news before you commit for years. These notes come from Replicate's pricing page and docs, and we did not run the platform ourselves.

Related pages

• Hosting hub and AI hosting overview
• Modal vs Runpod
• Private AI hub