Replicate Review 2026

What is Replicate?

Ten thousand images from FLUX Schnell cost $30 on Replicate. The same ten thousand from FLUX 1.1 Pro cost $400. Nothing about your account changes between those two numbers; only the model name in the API call does. That is the single most useful thing to understand in this replicate AI model hosting review and pricing breakdown, because Replicate has no monthly plans at all. You pay per image, per token or per second of GPU time, and the rate depends on what you run and on whether the model is public or your own.

Replicate is a San Francisco service that puts open and commercial models behind one HTTP API, with Python and Node clients, and lets you upload custom models packaged with its open-source tool, Cog. It is built for developers who want to ship an image, video or language feature this week without babysitting a GPU server. Figures below come from the pricing page, the billing and data docs, the terms and the Cog repository.

Replicate pricing per second: GPU rates and what an hour really costs

Per second billing applies to most public models and to anything you host yourself. The pricing page lists these single-card rates (INR figures are approx., converted at ≈₹94.7 per USD):

There is also a small 1-core CPU at $0.09 an hour. Multi-GPU machines cost a straight multiple of one card. Only two A100s ($10.08 an hour) or two L40S cards ($7.02) are open to any account. Every multi-H100 machine (2x at $10.98 up to 8x at $43.92 an hour), the H200 and larger A100 or L40S setups need a committed-spend contract.

Per-second rates look tiny until you multiply. A model that needs four seconds on an H100 costs about $0.006 a run. Run it 50,000 times a month and that is roughly $305 (about ₹28,900). That beats renting a server for bursty traffic, but often not for a steady load.

Replicate FLUX API cost per image, from Schnell to 1.1 Pro

Black Forest Labs' image models get heavy use (the FLUX.2 Pro page alone showed about 12 million runs in October 2026), and their official versions are priced by output rather than by seconds. From the pricing page:

• FLUX Schnell: $3 per thousand images, so about ₹0.28 an image (approx.).
• FLUX Dev: $0.025 an image, about ₹2.37 (approx.).
• FLUX 1.1 Pro: $0.04 an image, about ₹3.79 (approx.).

The newer second-generation models, including FLUX.2 Pro, are also live as official models, but the pricing table does not list them and we could not confirm a FLUX.2 Pro rate, so check the model page or your billing dashboard before you budget. Image prices for other vendors sit in the same range: Recraft V3 is $0.04 and Ideogram v3 Quality is $0.09. Video costs far more. Wan 2.1 image-to-video is $0.09 per second of 480p output and $0.25 per second at 720p, so a five-second 720p clip is $1.25.

Language models bill by tokens: DeepSeek R1 is $3.75 per million input tokens.

Calling FLUX from Python in five lines

The getting-started guide is short. Install the client with pip install replicate, create a token on the API tokens page, and export it as REPLICATE_API_TOKEN. Then call replicate.run with a model name such as black-forest-labs/flux-schnell and an input dictionary holding your prompt. Images come back as FileOutput objects whose .read() bytes you can save as a PNG, and language models can stream tokens.

For anything slower than a few seconds, skip the blocking call. Create the prediction, pass a webhook URL, and let Replicate POST the result to your server when it finishes. Webhooks matter for another reason too: API predictions, their inputs, outputs and logs are deleted after one hour by default. If your app does not copy the file somewhere within that hour, it is gone.

Cold boots, idle time and the private-model bill

Public models share a pool of GPU hardware. The billing docs say you pay only for active processing time on them; setup and idle time are free. The catch is speed. If nobody has called a model recently, your request waits while a container boots and loads weights. That cold start is free on public models, but the first response arrives noticeably later.

Private models and deployments flip the arrangement. You get dedicated instances and pay for every second they are online, including setup and idle. A minimum instance count keeps a model warm, but one T4 left running all month is about $591 (≈₹56,000), so set a max-instance cap too. One exception helps fine-tune users: fast-booting fine-tunes, marked in a model's version list, charge only for active time even when private.

The billing docs say a failed run is never charged. Cancelling a run on an official model may still cost money, and on private hardware or a deployment you pay for the time the instance was online, however the run ended.

Using Replicate's Cog to deploy a model you trained yourself

Cog is the open-source packaging tool behind every model on the site. It is released under the Apache 2.0 licence and has about 9.5k stars on GitHub. You describe the environment in cog.yaml (GPU on or off, system packages, Python version, a requirements file) and write a small Python class with a setup method that loads weights once and a run method that takes typed inputs. Cog turns that into a Docker image with an HTTP prediction server. The current README shows cog run for a local test, cog build to produce the image and cog serve to expose it on a port.

It needs Docker (plus the Buildx plugin if you use Docker Engine rather than Docker Desktop), and it runs on macOS, Linux and Windows 11 through WSL 2, not natively on Windows. Mac users can install it with Homebrew. The same container can go to Replicate or to your own servers, which reduces lock-in more than most hosted APIs allow.

Fine-tuning FLUX on your own product photos

Fine tuning on Replicate is mostly about LoRA adapters for image models. The docs point to replicate/fast-flux-trainer as the trainer for FLUX, and Stable Diffusion and SDXL LoRAs were supported early on. A D2C brand might train on twenty shots of its own bottle, then generate ad variations with the packaging intact. The trained weights become a model under your account, priced like any other private run or as a fast-booting fine-tune.

Remember the licence on the base model. Stable Diffusion 3.5 weights come under Stability AI's Community License, free for organisations under $1M in annual revenue, while SDXL uses the CreativeML Open RAIL++-M licence. Replicate's terms also say output rights are subject to each model's third-party terms.

Replicate vs Hugging Face Inference: same models, different bills

Hugging Face is the closest comparison, and the two are linked: Replicate is one of the providers Hugging Face's Inference Providers router can call. The difference is how you pay.

For an always-on dedicated GPU, Hugging Face is clearly cheaper on these list prices. Replicate wins on spiky traffic, because public models cost nothing while idle, and on its range of image and video models with a fixed per-output price.

Is a private model on Replicate safe?

A private model on a personal account is visible only to you, and models created inside an organisation can be shared with teammates. Deployments add private endpoints and audit logs of who accessed a model and changed its settings. The terms (updated 1 April 2026) say you keep ownership of your inputs, and Replicate passes you whatever rights it holds in the outputs. Replicate does get a licence to use your data to run the service and to train customer derivative models for you, and it can compile anonymised, aggregated usage data. Neither the privacy policy nor the terms we read mention a SOC 2 report, so enterprise buyers should ask sales directly. Predictions made in the web playground are kept until you delete them, unlike the one-hour API default.

Paying from India: dollars, Stripe and tax

There is no INR price list. Charges run through Stripe by credit card, debit card or bank transfer, and the terms add taxes on top of the fees. Usage is billed monthly in arrears or from prepaid credit. Under the terms, prepaid credit is non-refundable and each top-up expires at the end of the twelfth month after payment, so buy in small amounts. An Indian card needs international online payments switched on, and your bank's forex markup applies to every charge.

There is no fixed free tier either. Some models can be tried free, but billing is requested soon after and some features need it from day one.

What Cloudflare buying Replicate means for existing users

On 17 November 2025, Cloudflare and Replicate both announced that Replicate is joining Cloudflare. Replicate's own post promised that the API will not change, existing models keep working, and the company will carry on as a distinct brand. Cloudflare's post adds a plan to bring the whole Replicate catalogue to Workers AI. As of October 2026 we found no official notice that the deal has closed (Cloudflare's 2026 quarterly reports do not name Replicate), and the terms and privacy policy still name Replicate, LLC.

Replicate AI model hosting review and pricing: who should pay, who should look elsewhere

Pay for it if you are a small product team or agency that wants to run models with API calls, mixes several image and video models, and has uneven traffic. The per-output FLUX prices make cost per feature easy to forecast, and Cog gives you a clean exit path. Look elsewhere if you need a GPU busy around the clock: Hugging Face Inference Endpoints or a cloud GPU rental cost less per hour. Fal.ai and Together AI are the usual hosted alternatives, and Cloudflare Workers AI is worth watching, since Cloudflare plans to bring Replicate's models to it. Start with FLUX Schnell on a hundred prompts, check the bill, then decide whether a dedicated deployment is worth it.