Replicate says you can run open-source machine learning models with a cloud API, and also fine-tune or deploy your own. The homepage offers a Playground, a free start, and Node, Python, and HTTP examples. Public models are billed for processing time or for input and output. Private models, packaged with Cog, usually stay on dedicated hardware and are billed while they are online.
Official website · replicate.com
Use Replicate when the job is calling a hosted model, not when you want one image app with its own editor. Public models fit bursty calls because you only pay for the time a request is processed, or for a published output unit. Private models fit a queue you do not want to share, and the pricing page says most of them also bill setup and idle time. A homepage banner said Wan 3.0 is 30% off this week, so a listed rate can move. The pages do not state a monthly free credit.
The homepage says to run and fine-tune models, deploy custom models, all with one line of code, and to get started for free. Tabs on that page are Node, Python, and HTTP. The Node sample constructs a client with an auth token and calls run with an input object that includes a prompt. A separate line says you can compare models in the Playground. Named outputs include images, speech, music, restored images, videos from images, captions, and language models. The docs index adds Colab, an image fine-tune, a custom deploy, and webhooks. Community models sit next to proprietary ones. Every model’s input fields are not listed on these pages.
Official documentationYou only pay for what you use on Replicate. Some models are billed by time, others by input and output. On public models, most are billed by the time they take to run, and the price-per-second varies according to the hardware in use. When you run or train one of those models, you only pay for the time it takes to process your request. Other public models publish an output unit. Examples include flux-1.1-pro at $0.04 per output image, flux-dev at $0.025 per output image, and flux-schnell at $3.00 per thousand output images. DeepSeek R1 is $0.01 per thousand output tokens and $3.75 per million input tokens. Ideogram v3 quality is $0.09 per output image. Each model page is said to carry its own estimate.
Official documentationPrivate models are not limited to the public catalog. You can deploy your own custom models using Cog, described as an open-source tool for packaging machine learning models. Unlike public models, most private models run on dedicated hardware so you do not share a queue. You pay for all the time instances are online: setup, idle waiting, and active processing. If traffic spikes, the page says capacity scales up and down. Fast booting fine-tunes are billed while active, not for idle time, and they are labeled in the version list. An Nvidia A100 (80GB) example is $0.001400 per second and $5.04 per hour. Extra multi-GPU capacity can require a committed spend contract, and that contract price was not verified here.
Official documentationFLUX.2 flex documents its own controls. You adjust how many steps to run and how strongly the model should follow your prompt. The page aims it at typography-heavy layouts and at edits that use several reference photos. Clean type, readable captions, and complex layouts are the stated strength. The multi-reference feature lets you blend up to ten images into a single output. Guidance around 2-3 is freer. Values of 4-5 follow the prompt more literally, and the page suggests starting near 3.5. About 6-10 steps are the quick draft. About 20-50 steps are slower and described as sharper. Negative prompts are said not to work. This is the model page’s guidance, not a generation run here.
Official documentationIs to compare two image models in the Playground before writing a client. Read each price unit. A per-image model and a per-second model can bill the same prompt differently. Then use the Node, Python, or HTTP path on the homepage. Is a reading of the docs.
For a custom checkpoint is Cog, then a private deployment. Most private models bill setup, idle, and active time. Turn the deployment down when the traffic is gone. If the version list labels a fast booting fine-tune, idle time is not billed.
Before a large batch is to multiply the published unit. One thousand flux-schnell images are listed at $3.00. Flux-dev at $0.025 per image is a different order of cost. DeepSeek mixes $0.01 per thousand output tokens with $3.75 per million input tokens. Add an A100 at $5.04 per hour only if the model must stay online. Ordinary private models also bill idle time. This suggested arithmetic uses the published rates only.
Open the Playground from the homepage and pick a public model whose output matches the job.
Use the free start, then choose the Node, Python, or HTTP example. The Node sample sends an auth token and an input prompt.
Check whether that model bills hardware time or an output unit, and whether a private deploy would also bill idle time.
Call run for a public model, follow the fine-tune guide for an image model, or package a custom model with Cog.
There is no single subscription price on the public pages. You only pay for what you use on Replicate. Public examples include $0.04 per output image, $0.025 per output image, $3.00 per thousand output images, $0.01 per thousand output tokens, $3.75 per million input tokens, and $0.09 per output image. Private hardware is listed per second and per hour. The homepage says to get started for free and does not publish a monthly credit.
Calls against community or proprietary models on shared hardware.
$0.04.
Example per output image; other public models use time or tokens.
Custom Cog models that need dedicated hardware.
$5.04.
Example per hour for Nvidia A100 (80GB); also listed per second.
The $0.04 figure is the flux-1.1-pro example, not a platform minimum. Flux-dev, flux-schnell, DeepSeek R1, and Ideogram v3 quality use the other rates above. The A100 line is $0.001400 per second and $5.04 per hour, and other GPUs on the same page differ. Extra multi-GPU capacity is described as available with committed spend, without a published contract amount. Fast booting fine-tunes do not follow the idle rule. Rates can change, including the temporary Wan discount on the homepage.
The job is routing language-model calls, not image or custom GPU deploys.
OpenRouter is its own developer-tool listing. Replicate’s public pages include language models and also image, speech, music, and video runs, plus private hardware.
You want Ideogram’s own product rather than a hosted model call.
Replicate’s pricing page lists an Ideogram v3 quality model at $0.09 per output image inside a multi-model API. Pick Ideogram when you want that product’s own app, and Replicate when Ideogram is one model among many calls.
The homepage says to try for free and to get started for free. Pricing says you only pay for what you use, by time or by input and output. No monthly free credit, free prediction count, or free hardware hour is stated. Read the model meter before a batch.
SourceMost public models are billed by run time. The price-per-second follows the hardware, and you pay for processing time when you run or train. Some models bill input and output instead. Examples include $0.04 and $0.025 per image, $3.00 per thousand images, DeepSeek token rates, and $0.09 per Ideogram image.
SourcePackage a custom model with Cog. Most private models use dedicated hardware and bill setup, idle, and active time. Fast booting fine-tunes bill only while active and are labeled in the version list. An A100 example is $0.001400 per second and $5.04 per hour. Extra capacity can require committed spend, with no contract amount published.
SourceThe homepage lists images, speech, music, restored images, videos from images, captions, and language models. Compare them in the Playground or call Node, Python, or HTTP. Docs also list Colab, image fine-tuning, custom deployment, and webhooks. Inputs depend on the model. Flex takes a prompt, steps, guidance, and up to ten reference images.
SourceFlex exposes steps and prompt strength. Guidance around 2-3 is freer. Values of 4-5 follow the prompt more literally, with a suggested start near 3.5. About 6-10 steps are the quick draft. About 20-50 are slower and described as sharper. Up to ten reference images can be blended. Negative prompts are said not to work.
SourceOct 6, 2026 · NavHunt editorial
NavHunt did not run Replicate, did not call the API, and did not generate an image in the Playground. This profile restates the homepage, pricing page, docs index, and FLUX.2 flex model page reviewed on 6 October 2026. Latency and actual charges were not measured. This description is based on Replicate's official materials reviewed on 2026-10-06, including Replicate - Run AI with an API, Pricing – Replicate, Documentation – Replicate, FLUX.2 [flex] model page.
Start from the free entry point, then read the model’s price unit before a large batch.