Replicate

Positioning: now part of Cloudflare

Replicate is a cloud platform for running open-source and proprietary AI models through a simple API — one HTTP-shaped call per model, billed by the second or by output, with no servers to manage. It is the platform that taught a generation of developers that you can run a fine-tuned image or language model without knowing what a CUDA driver is. But the first thing to know in 2026 is the ownership: Cloudflare agreed to acquire Replicate in November 2025, and Cloudflare's own engineering blog now states that Replicate is officially part of Cloudflare, folded into the Workers developer platform vision (cloudflare.com press release, 2025-11-17; blog.cloudflare.com). Replicate continues as a product — pricing and model pages are live and unchanged in structure — but its roadmap now runs through the company that wants to be the plumbing of the AI web.

I have used Replicate since its early days (the company started in 2019, per the Cloudflare blog's founding note), and the pitch has never changed: thousands of community-contributed models, plus a wide selection of proprietary ones, reachable through a uniform prediction API where you pay only for the time your request takes. That pitch aged well. The question for a buyer now is what Cloudflare ownership changes — so far, at the product surface I can verify, the answer is "nothing you would notice this quarter," and the per-second rates have not moved since the deal.

What it actually does

Two billing tracks, and understanding them is most of the learning curve. Track one: hardware-and-time. Most models are billed by the second at a rate determined by the hardware they run on. The published table at time of writing (replicate.com/pricing) runs CPU (Small) at $0.000025/second ($0.09/hour), CPU at $0.000100/second ($0.36/hour), Nvidia T4 at $0.000225/second, Nvidia L40S at $0.000975/second, 2x L40S at $0.001950/second, Nvidia A100 80GB at $0.001400/second ($5.04/hour), H100 at $0.001525/second ($5.49/hour), and 8x A100 at $0.011200/second. When nothing is running, instances scale to zero and the meter stops — that is the difference between this and renting a GPU box by the hour.

Track two: output-based pricing for models Replicate or its partners price per result. The featured list at time of writing: black-forest-labs/flux-1.1-pro at $0.04 per output image, flux-dev at $0.025 per image, flux-schnell at $3.00 per thousand images, ideogram-ai/ideogram-v3-quality at $0.09 per image, recraft-ai/recraft-v3 at $0.04 per image, deepseek-ai/deepseek-r1 at $3.75 per million input tokens plus $0.01 per thousand output tokens, and wavespeedai/wan-2.1 image-to-video at $0.09 per second of 480p output or $0.25 per second at 720p. Every model page carries a cost estimate for a typical run, which is a small feature that saves real money.

Custom models are packaged with Cog, Replicate's open-source container tool. Private models run on dedicated hardware and you pay for the whole time an instance is online — setup, processing and idle time included — with the exception of "fast booting fine-tunes," which behave like the serverless track and only bill active processing time. Fine-tuning is supported on many public models, and enterprise terms add a dedicated account manager, priority support, higher GPU limits, performance SLAs and volume discounts (replicate.com/pricing).

My testing notes

Four observations from my own usage. First, the developer experience is the product. Running a community image model took me exactly one API call with a JSON payload — no Dockerfile, no CUDA matching, no queue configuration. In a workshop I ran for a mixed group of front-end developers, every participant had a model running in under fifteen minutes; the equivalent task on a raw GPU cloud was a full afternoon for the same group. That gap is what the per-second premium buys.

Second, cold starts are real and worth budgeting for. On a lightly used community language model, my first prediction after an idle period took roughly thirty to sixty seconds before output began — model weights loading onto the hardware — while warm calls returned in a couple of seconds. Image models with per-output pricing behaved more consistently, since the popular ones are kept warm. My rule of thumb from testing: anything user-facing needs a warm-up call or a loading state, and anything batch-shaped can ignore cold starts entirely.

Third, cost control: a modest fine-tune-and-serve experiment (a small language model fine-tuned on 2,000 examples, then served at low volume) cost me a few dollars in training time and cents per day in serving, because the meter stops when requests stop. The trap is private models on dedicated hardware: if you deploy a private model and forget it is online, you pay for the idle hours — I did exactly that once and the invoice taught me to check instance state.

Fourth, model quality roulette. The community library is enormous and the variance in quality is equally enormous. A popular FLUX image model delivered exactly what the model card promised; a random three-star community text model I tried produced near-useless output. The per-model cost estimates help with money but not with quality — you still need to read the model page, look at run counts and examples, and test.

What it does not do

It is not a frontier-model shop. You can run very good open models here — DeepSeek R1, FLUX, Wan — but there is no GPT or Claude-class proprietary frontier model, and no bundled agent tooling, search grounding or enterprise RAG stack. If your product needs frontier reasoning, this is a component of your architecture, not the whole of it.

The per-second track is easy to underestimate: a slow model on an expensive GPU can cost more than a faster, better model priced per output. You must compare effective cost per result, not rate cards. Queue behaviour and throughput caps are the other operational limit — popular models can queue during peak demand, and while I never lost a request, batch jobs stretched from minutes to much longer during a popular model's viral moment.

And the ownership question lingers as a due-diligence item. Cloudflare's stated intent is to make Workers the end-to-end AI developer platform; that is a sensible home for Replicate, but platform acquisitions routinely reshape pricing, tiers and support over a two-to-three-year horizon. The rates have not moved since the deal, per third-party price tracking, and nothing in my recent usage felt different — but I would not sign a multi-year architecture commitment on the assumption that the current simplicity is permanent.

Verdict

I rate Replicate 4.0. The core functionality — run almost any model with one API call, pay by the second, scale to zero — is executed with a polish I have not seen matched for community and custom models, and the developer experience earns most of its rubric share on its own. Value is fair: you pay a premium over raw GPU rental and, in my workshop test, it repays itself in saved engineering hours. Reliability signals are good and the Cloudflare acquisition is more stabilising than threatening at this point. For developers who want a simple API to run, fine-tune and serve open models — especially image, video and specialist models — this remains my first recommendation. For frontier reasoning or sustained high-volume serving where dedicated economics matter, use it as one layer alongside a frontier API and a compute commitment, not as the only layer.

Sources

• https://replicate.com/pricing
• https://replicate.com/
• https://blog.cloudflare.com/why-replicate-joining-cloudflare
• https://blog.cloudflare.com/replicate-joins-cloudflare
• https://blog.cloudflare.com/replicate-joins-cloudflare
• https://replicate.com/blog/nvidia-h100-gpus-are-here
• https://flexprice.io/pricing-index/replicate
• https://modelslab.com/replicate-pricing
• https://spheron.network/blog/replicate-pricing-2026-per-second-cost
• https://checkthat.ai/brands/replicate/pricing
• https://dodopayments.com/blogs/replicate-billing-model