Skip to main content
nexeai

Replicate: our review

Replicate runs open AI models behind an API, billed by the second of compute rather than by subscription. It's a platform for developers, not a tool you open and use without writing code — and since late 2025, it belongs to Cloudflare.

Paid
Go to Replicate's website (opens in a new tab)

What we take away from it

What counts in its favor

  • Billing by the actual second of compute used (0.000225 $/s on a T4 GPU, up to 0.001525 $/s on an H100), with no monthly subscription or minimum: a public model costs nothing as long as no request is running.
  • A single API call gives access to a catalog of more than 50,000 open models, with no need to set up or maintain the GPU infrastructure behind it.
  • The acquisition by Cloudflare, announced on November 17, 2025, promises an unchanged API and integration with the Cloudflare network: Replicate itself writes that 'the API isn't changing' and that models already in use keep running.

What counts against it

  • Nothing can be used without writing code: HTTP API, Python client or Node.js client are the only documented entry points. An SMB with no developer, in-house or contracted, can't get anything out of it directly.
  • A private model also bills idle time — startup and waiting included — not just active compute, unlike public models: a misconfiguration can drive up the bill without any inference having taken place.
  • The terms of use grant Replicate a broad license over submitted data (use, reproduction, creation of derivative works) to provide the service and train the customer's custom models; the privacy policy doesn't say anywhere whether inputs are also used for anything else.

A good choice for a technical team that wants to test or deploy an open model without managing a GPU. With no in-house developer or contractor, the tool stays out of reach.

The facts

Data hosting
Hosting not verified
Data verified on
30 July 2026

What happens to your data

Is your data used to train their models?
Subprocessors
The terms of use grant Replicate a license to use customer data to provide the result (Output) and to 'train and generate Customer Derivative Models' — models obtained by fine-tuning on the customer's own data, distinct from training Replicate's general models. The privacy policy does not specify where servers are located.

Replicate in detail

Replicate runs open AI models behind an API, without anyone having to rent or maintain a GPU server. The principle: a model from the catalogue — image, voice or video generation, or a large language model — gets called in a single line of code, runs for the duration of a request, and the bill stops when the computation does.

A tool for developers, not a marketing or accounting team

Nothing on Replicate can be used without writing code. The documentation offers only three entry points: an HTTP API, a Python client, a Node.js client — plus examples for Google Colab. There is no interface where a non-developer could drop a file and get back a result, unlike a consumer tool. An SME or a firm without a developer, in-house or through a service provider, has no use for Replicate: it's a soldering iron for someone who already knows how to solder, not a finished tool.

The bill follows the computation, not the month

The pricing comes down to one sentence from Replicate itself: "you only pay for what you use." Computation is billed by the second, on a scale that runs from $0.000025 for CPU to $0.001525 for an H100 GPU, and some models publish a rate per token or per image rather than per compute time. For a public model, the wait between two requests is free. That stops being true once you host a private or fine-tuned model on dedicated hardware: start-up time and idle time are billed too, which can catch out a team that isn't monitoring its usage.

What the Cloudflare acquisition changes

Cloudflare announced on 17 November 2025 its intention to acquire Replicate, to give its own developers direct access to the catalogue's more than 50,000 models. Replicate writes on its own blog that the brand keeps existing, that the API doesn't change, and that any project already built on it keeps running. Still, the infrastructure will gradually integrate with Cloudflare's: a point to watch for anyone planning to host a model there over several years.

Where it falls short

Two reservations. The first: the terms of service grant Replicate a broad licence over submitted data, including for training models derived from the customer — a fine-tuning requested and paid for by the customer themselves, but the privacy policy doesn't specify anywhere whether the same data is used for anything else. The second: per-second billing, with no minimum or predictable tier, complicates a fixed monthly budget — fine for a prototype, less so for an SME that wants a stable figure to present to a management committee.

Our recommendation

A solid choice for a technical team that wants to quickly test an open model without buying a GPU, provided it has a developer able to integrate the API and monitor the bill. See also the other model platforms in the comparison. For business automation without code, it's better to turn to our automation services or an AI agent built on top, rather than to Replicate directly.

Who it's for, and who it isn't

Suited for

  • A team with a developer able to integrate an API, who wants to try or deploy an open model (image, audio, video, LLM) without buying or renting a GPU
  • A startup or agency that needs to prototype quickly across several models before choosing which one to host permanently elsewhere

Not a fit

  • An SMB or independent professional with no developer, in-house or contracted: there is no ready-to-use interface, only an API
  • An organization that needs a predictable compute budget from month to month: per-second billing and the cost of idle time on private models make the bill hard to forecast

Frequently asked questions

Do you need to know how to code to use Replicate?
Yes, always. The only documented entry points are an HTTP API and client libraries in Python and JavaScript: with no developer, in-house or contracted, the tool cannot be used.
How does Replicate bill compute?
By the second of hardware used, from CPU to H100 GPU, with no subscription or minimum for public models. Some models bill by token or by generated image instead; a private model also bills its idle time between two requests.
Is data I send to Replicate used to train its models?
The terms of use allow Replicate to use it to deliver the requested result and to train the customer's own custom (fine-tuned) models. Nothing in the public pages specifies whether that data also feeds the platform's general models.
Does Replicate still belong to its founders?
No: Cloudflare announced its acquisition of Replicate on November 17, 2025. Replicate states it keeps existing as a distinct brand, with the same API, while gradually integrating with Cloudflare's infrastructure.
What's the difference with Hugging Face?
Hugging Face is first a model repository, with managed hosting (Inference Endpoints) as a paid option. Replicate is built from the ground up as an on-demand compute API, with no equivalent community repository to self-host.

Alternatives to Replicate

Our sources

Prices and hosting terms change often. We recheck them periodically, but the publisher's page remains the only up-to-date reference.

Is Replicate the right choice for you?

We measure this tool against your real constraints — budget, privacy, your team's skill level — before you sign an annual subscription.

Book a call