Replicate Pricing Explained: Plans, Costs and What You Get

Replicate has no monthly plans: you pay per second of hardware time or per output. This guide explains the GPU rates, billing rules, and what jobs really cost.

By Galactic News Editorial · Published 2026-11-10T01:25:53+00:00 · Updated 2026-11-10T01:25:53+00:00

Replicate Pricing Explained: Plans, Costs and What You Get

Replicate has no monthly plans. You pay for what you use, from prepaid credit. Public models bill per second of hardware time or per output, depending on the model. Private models run on dedicated hardware and bill for all the time they are online. This guide explains the rates, the billing rules, and what a typical job costs.

Replicate Pricing: The Hardware Table

For models billed by time, the price depends only on the hardware the model runs on. The model itself does not change the rate. A small model on a T4 costs the same per second as a large model on a T4. The pricing page lists these rates.

Multi-GPU configurations above 2x and H200 capacity are available through committed-spend contracts. The pricing page lists the rates but notes the capacity terms. Check the page for the current list before you plan a large deployment.

How Billing Actually Works

Replicate bills public models only for the time the model spends processing your request. If a model runs for 30 seconds, you pay for 30 seconds. There is no charge for queue time and no idle charge on public models. That makes public models cheap for spiky traffic and prototypes.

Private models work differently. Most private models run on dedicated hardware, so you do not share a queue with anyone else. You pay for all the time the instance is online: setup time, idle time waiting for requests, and active processing time. If traffic is heavy, Replicate scales the deployment up and down automatically.

That idle billing is the catch to watch. A private H100 left warm and unused costs $5.49 an hour, every hour. Over a month that is about $4,000. Set up auto-scaling and scale-to-zero where the deployment supports it. Turn off deployments you are not using.

There is one exception. Fast-booting fine-tunes are billed only for the time the model is active and processing requests. You do not pay for idle time, as with public models. Fast-booting fine-tunes are labelled as such in the model's version list.

Per-Output Pricing

Many popular models do not bill by the second at all. They charge a fixed price per output. Image models charge per image. Video models charge per second of video. Some language models charge per token. The rate is set per model and shown on that model's page.

Per-output pricing is the easiest kind to forecast. You know the cost of one image before you generate it. Multiply by the number of outputs and you have the bill. That predictability is why production teams often prefer per-output models for user-facing features.

Time billing suits a different pattern. It fits models with variable run times, custom models, and batch jobs where the output count is not the natural unit. If you run your own model through Cog, Replicate's open-source packaging tool, time billing applies.

Every model page shows its own price or a cost estimate. Check the page before you build. Rates can change, and a model that was cheap last month may have moved to a larger hardware tier.

What Replicate Costs in Practice

Hourly rates make the math easy. One hour of continuous H100 time costs $5.49. One hour on an A100 costs $5.04. One hour on a T4 costs $0.81. A full day of continuous H100 time costs about $131.76.

Real jobs are shorter than that. A ten-minute run on a T4 costs about $0.14. The same ten minutes on an H100 costs about $0.92. Short, bursty jobs are where Replicate is cheapest, because you pay only for active seconds and there is no monthly fee to cover.

The expensive mistake is the forgotten deployment. A private model left running on a dedicated H100 for a month costs roughly $4,000, even if it serves no requests. A public model with the same usage costs only the seconds it actually processed. Keep private deployments for steady traffic, and use public models for everything else.

There is no free tier on the pricing page. You pay from the first run, so budget a small test amount when you try a new model. A few dollars of prepaid credit is enough to learn a model's real cost per request.

Who Replicate Fits Best

Replicate hosts thousands of open-source models contributed by the community, and you can deploy your own custom models with Cog. That combination is the core of the product: one API for many models, with billing tied to hardware time or output.

Bottom Line

Replicate pricing is simple and honest. No monthly plans, no minimum spend. Public models cost $0.09 to $5.49 an hour depending on hardware, and you pay only for processing time. Private models cost the same per second but bill for idle time too, so watch deployments you leave running. Start with public models, check each model page for per-output rates, and scale to dedicated hardware only when traffic justifies it.