What it is
Replicate lets developers run AI models with one API call, without managing servers. Its catalogue is strongest for image, video and audio models such as FLUX, Ideogram and Wan, and you can package and deploy your own model with its open-source tool, Cog.
Cloudflare completed its purchase of Replicate on 1 December 2025. Replicate still runs under its own name with the same API, and in April 2026 Cloudflare said the team had merged into its AI Platform group and was moving Replicate's hosted models onto Cloudflare infrastructure and into AI Gateway.
Billing is pay-as-you-go. Many popular models charge per output, for example $0.04 per image for FLUX 1.1 Pro. Others bill by the second for the hardware they run on, from $0.09 an hour for a small CPU to $5.49 an hour for an NVIDIA H100.
Our review
Replicate's value is breadth behind one interface. Each model has a web page, a playground and copy-paste code, and the API is the same whether you call a FLUX image model, a Wan video model or a speech model. New media models tend to appear there early, and many official models bill per output, so the cost of an image or a second of video is known before you ship.
Custom models are a different deal. Anything you deploy with Cog runs on dedicated hardware billed for every second it is online, including boot and idle time, according to the pricing page; only fast-booting fine-tunes escape the idle charge. One L40S left running all month costs about $2,560 and one H100 about $4,000, more than many specialist GPU clouds charge for the same card. Reviewers also flag cold starts, the delay while a little-used model loads onto a GPU.
The data defaults are friendlier than most. The docs say inputs, outputs, files and logs from API predictions are deleted after an hour unless you copy them out, which suits privacy-sensitive work but surprises teams who expected Replicate to host their output URLs. Predictions made in the web playground are kept indefinitely. Replicate does not publish a SOC 2 report or a training-use policy on its site, so enterprise buyers have to ask.
Ownership is the open question. Cloudflare paid $57.4 million, according to its SEC filing, and its April 2026 platform post said Replicate's models were being replatformed onto Cloudflare infrastructure and brought into AI Gateway, which already lists Replicate as a provider. So far the API, catalogue and pricing are unchanged; teams building on Replicate today are effectively betting on how much of it survives as a separate product.
Where it earns its place
A two-person team building a product-photo app
Generate and retouch images on demand without owning GPUs
Call FLUX 1.1 Pro at $0.04 an image; 50,000 images a month costs $2,000 with no idle charge. Save outputs to your own storage, because API results are deleted after an hour.
An ML engineer serving a fine-tuned video model
Expose a custom model as an API for an internal tool
Package it with Cog and deploy on an L40S at $3.51 an hour; kept warm around the clock that is about $2,560 a month, billed for idle time too. Scaling to zero cuts cost but brings cold starts back.
A product manager at a marketing agency
Compare several video models before committing to one vendor
Run the same brief through Wan and other video models in the playground, where Wan 2.1 image-to-video costs $0.09 a second at 480p and $0.25 at 720p. Playground predictions are kept indefinitely, so delete client material afterwards.
Before you commit
- Setup
- Sign up, add a payment method and an API token, then paste the model page's Python, Node or HTTP snippet; accounts without a payment method are limited to 6 requests a minute.
- Integrations
- Official Python, Node.js, Go and Swift clients, a hosted MCP server, a REST API with webhooks, Cog for packaging custom models, and Cloudflare AI Gateway as a provider.
- Data and privacy
- API prediction inputs, outputs and logs are deleted after an hour by default; web predictions are kept indefinitely; SOC 2 status and training-use terms are not published on Replicate's site.
- Cost at scale
- A private model kept warm on one H100 costs about $4,000 a month and on 8x H100 about $32,000; per-output models scale linearly, so 1 million FLUX Schnell images cost $3,000.
- Limits
- 600 prediction requests a minute and 3,000 for other endpoints by default; stricter limits when credit runs low; 4x and 8x A100, multi-H100 and H200 capacity needs a committed-spend contract.
What's good
- One API across thousands of image, video, audio and language models
- Per-output pricing on official models makes unit costs predictable
- API prediction data deleted after one hour by default
- Cog packages and deploys your own models
- Reachable through Cloudflare AI Gateway
What's not
- Custom models bill for boot and idle time
- H100 at $5.49 an hour is dear next to specialist GPU clouds
- Cold starts on less-used models
- No published SOC 2 report or training-use policy
- Long-term product shape depends on Cloudflare's integration plans
Best for
Developers adding image, video or audio generation to an app without running their own GPUs.
Alternatives
Replicate: common questions
- Is Replicate worth it?
- Replicate is where a developer goes to try a new image or video model this afternoon, and where the bill starts to hurt once one model runs all day. We rate it 4.2 out of 5.
- How much does Replicate cost?
- Pay per output (e.g. FLUX 1.1 Pro $0.04/image, FLUX Schnell $3 per 1,000 images) or per second of hardware: T4 $0.81/hour, L40S $3.51/hour, A100 80GB $5.04/hour, H100 $5.49/hour, 8x H100 $43.92/hour. No free tier listed; volume discounts for large spend.
- What are the downsides of Replicate?
- Custom models bill for boot and idle time; h100 at $5.49 an hour is dear next to specialist GPU clouds; cold starts on less-used models; no published SOC 2 report or training-use policy; long-term product shape depends on Cloudflare's integration plans.
- Who is Replicate best for?
- Developers adding image, video or audio generation to an app without running their own GPUs.
- How much does Replicate cost for a team or in production?
- A private model kept warm on one H100 costs about $4,000 a month and on 8x H100 about $32,000; per-output models scale linearly, so 1 million FLUX Schnell images cost $3,000.
- Is Replicate safe for company data?
- API prediction inputs, outputs and logs are deleted after an hour by default; web predictions are kept indefinitely; SOC 2 status and training-use terms are not published on Replicate's site.
- What are the best Replicate alternatives?
- The closest alternatives we have reviewed are fal, Hugging Face, Modal and Together AI.
Answers checked Oct 22, 2026.
Make Replicate? Show this rating on your site
