Replicate
API to run hosted image, video and audio models, billed per second of GPU time or per output
Replicate is a hosted API for running AI models: open-weights image, video and audio models plus hosted models from Google, OpenAI, ByteDance, Alibaba and others. Billing is per second of GPU time, or per output for some models, with no subscription tier on the pricing page. Cloudflare agreed to acquire it in November 2025; it still runs under its own brand and existing APIs.
- 3 creators using
- 5 news items tracked
- 2021 first launched
- Paid pricing
- Moderate learning curve
What Replicate can do
- Infinite canvas No
- AI agent No
- Node / flow builder Not confirmed
- Public API Yes
- MCP server Yes
12 of 26 capabilities confirmed
| Capability | Supported | Details |
|---|---|---|
| Text to image | Yes | Hosted image models, e.g. google/nano-banana-2, openai/gpt-image-2, black-forest-labs/flux-2-pro |
| Image editing | Not confirmed | |
| Text to video | Yes | alibaba/wan-3, bytedance/seedance-2.5, xai/grok-imagine-video, prunaai/p-video-2-pro |
| Image to video | Yes | Wan 2.1 image-to-video models listed with per-second prices |
| Video to video | Not confirmed | |
| Native audio in video | Yes | bytedance/seedance-2.5 generates audio and video together (can be turned off) |
| Lip-sync | Not confirmed | |
| Voice / TTS | Yes | Speech models such as inworld/realtime-tts-2; voice cloning not verified |
| Music & SFX | Yes | minimax/music-2.6 |
| Upscaling | Not confirmed | |
| Character consistency | Not confirmed | |
| Camera control | Not confirmed | |
| Custom training | Yes | Fine-tune open models such as SDXL on your own data; run them via the API |
| Third-party models | Yes | Hosts third-party models from Google, OpenAI, ByteDance, Alibaba, xAI, Black Forest Labs, Krea and Recraft |
| Capability | Supported | Details |
|---|---|---|
| Infinite canvas | No | |
| Node / flow builder | Not confirmed | |
| AI agent | No | |
| Chat assistant | No | |
| Team collaboration | Not confirmed |
| Capability | Supported | Details |
|---|---|---|
| Public API | Yes | REST API with Python and JavaScript SDKs; async predictions and webhooks |
| MCP server | Yes | Hosted MCP server (mcp.replicate.com) and replicate-mcp npm package |
| Mobile app | Not confirmed | |
| Desktop / local | Partial | Cog packages your own model for self-hosting; the platform itself is cloud-only |
| Open source | Partial | Cog is open source; many hosted models are open-weights, others are closed (e.g. Google, OpenAI models) |
| Capability | Supported | Details |
|---|---|---|
| Free tier | Not confirmed | |
| Commercial use | Not confirmed |
Yes Partial No Not confirmed by the vendor
What makes Replicate different
- 01
Thousands of open and hosted models behind one REST API and SDKs (Python, Node), with no GPU setup or CUDA drivers
- 02
Billing is per second of hardware time from $0.000025/s (CPU) to $0.001525/s (H100), or per output image or per second of video for models priced that way
- 03
Hosted MCP server at mcp.replicate.com lets Claude or Cursor search, compare and run models directly
- 04
Cog is the open-source packaging tool for putting your own model behind the same API, with autoscaling to zero
- 05
Fine-tunes on fast-booting hardware are billed only for active time; dedicated deployments bill for setup and idle time too
Latest from Replicate
- Apr 15 How to make remarkable videos with Seedance 2.0 Seedance 2.0 arrives on Replicate, offering high-fidelity AI video generation with improved temporal consistency and motion control for creators.
- Feb 24 How to prompt Seedream 5.0 Replicate's Seedream 5.0 introduces multi-step reasoning and example-based editing to improve accuracy in AI image generation.
- Feb 18 Recraft V4: image generation with design taste Recraft V4 launches on Replicate, offering art-directed image generation, precise text rendering, and the ability to export editable SVGs for graphic design workflows.
- Nov 26 Run Isaac 0.1 on Replicate Replicate now hosts Isaac 0.1, a lightweight vision-language model designed for precise spatial reasoning and real-world object perception.
- Nov 25 Run FLUX.2 on Replicate Replicate now supports FLUX.2, offering creators a high-performance environment to run the latest professional-grade image generation and editing models with multi-reference support.
About Replicate
Replicate lets you call a model over HTTP without managing GPUs. You pick a model, send inputs, and get back an output file or JSON, with logs and metrics on each prediction. Its public catalog covers image models such as google/nano-banana-2, openai/gpt-image-2, black-forest-labs/flux-2-pro and krea/krea-2-medium, video models such as bytedance/seedance-2.5 (up to 30 seconds, with audio generated in the same pass), alibaba/wan-3 and xai/grok-imagine-video, plus speech and music models. You can also fine-tune an open model such as SDXL on your own images and run the result through the same API.
What changed in 2026: the Cloudflare acquisition (announced November 2025) is the big one. Cloudflare says existing APIs and models keep working, and it plans to add Replicate's catalog and fine-tuning to Workers AI and route it through AI Gateway for cost tracking. Smaller 2026 product changes were an MCP server listing in the official MCP Registry (February), a fallback model for Nano Banana Pro when Google's API is at capacity (March), and agent skill files for coding assistants (April). The public changelog shows nothing after 21 April 2026, so some later releases may not be visible from outside.
For a film or video team, Replicate is a developer tool, not a creative app. It fits when you are building a pipeline: batch-generating frames, running upscaling or transcription overnight, or putting a model behind an internal tool or a web app. A filmmaker who only wants to prompt a video model will get more from a finished interface. The per-second GPU rates are cheap for short jobs, but per-second video models add up quickly on long clips.
What's new in Replicate
-
Agent skills for Replicate
Markdown skill files teach coding assistants how to run models through the Replicate API; installable with npx skills add replicate/skills.
Source: replicate.com ↗ -
Fallback model for Nano Banana Pro
With allow_fallback_model set to true, Nano Banana Pro switches to Seedream 5.0 lite when Google's API is at capacity; off by default.
Source: replicate.com ↗
Models available in Replicate
-
Seedance 2.5 Latest $$$ ByteDance Native audio, 4 to 30 seconds per generation or auto duration
-
Wan 3.0 Latest $ Alibaba alibaba/wan-3; Wan 2.1 i2v billed at $0.09/s (480p) and $0.25/s (720p)
-
Grok Imagine Video $$ xAI xai/grok-imagine-video
-
P-Video 2 Pro Latest — Pruna AI prunaai/p-video-2-pro
-
Nano Banana 2 $$$ Google google/nano-banana-2; Nano Banana Pro has an optional fallback model
-
GPT Image 2 — OpenAI openai/gpt-image-2
-
FLUX.2 [pro] $$ Black Forest Labs black-forest-labs/flux-2-pro; also flux-2-max and flux-2-flex listed
-
Recraft V4 Styles Pro SVG $$ Recraft recraft-ai/recraft-v4-styles-pro-svg
-
Krea 2 Medium Latest — Krea krea/krea-2-medium
-
Realtime TTS 2 Latest — Inworld inworld/realtime-tts-2 (speech)
-
Music 2.6 — MiniMax minimax/music-2.6
$ · $$ · $$$ = relative cost against other models of the same type, on fal.ai list prices. Cost inside Replicate depends on its own plans and credits.
Key Features
- REST API with Python and JavaScript SDKs; predictions run async with webhooks
- Thousands of open-source models plus hosted models from Google, OpenAI, ByteDance, Alibaba, xAI and Black Forest Labs
- Per-second GPU billing from CPU to 8x H100, plus per-output pricing on some public models
- Fine-tune open models such as SDXL on your own images and run them through the same API
- Cog open-source tool to package and deploy your own models
- Hosted MCP server for Claude and Cursor, and an experimental local code mode
- Prediction logs and performance metrics per run
- Batch processing: run thousands of predictions in parallel
- Nano Banana Pro fallback model option when the primary API is at capacity (March 2026)
- Agent skill files for coding assistants (April 2026)
When to reach for it — and when to skip
Reach for it when…
- No GPU, CUDA or driver setup; one API call runs a model
- Cheap for short jobs: an H100 second costs about $0.0015
- Wide model catalog, including newer video models with native audio (Seedance 2.5)
- MCP server makes it easy to run models from a chat session
- Own models can be deployed with Cog and scale to zero when idle
Skip it when…
- Developer tool: no timeline, canvas or prompt UI for editors
- Per-second video models add up fast on long clips and many takes
- Cold starts on less-used models are slow; third-party reviews cite roughly 10 to 30 seconds
- Output licensing depends on each model's own terms, so commercial rights must be checked per model
- Public changelog shows nothing after 21 April 2026, so recent releases are hard to confirm
Best For
✓ Ideal for
- Batch-generating stills or frames overnight, such as concept art or look-dev variations
- Adding image, video or speech generation to an internal tool or a web app
- Transcribing rushes or interviews in bulk and posting results to a webhook
- Upscaling or converting large sets of clips from a script, without local GPU VRAM limits
- Testing many models quickly through one API before committing to one vendor
✗ Not built for
- Editors and directors who want a creative interface, timeline or canvas
- Latency-critical live features that cannot tolerate cold starts
- Long-running video jobs on a tight budget, where per-second billing is hard to predict
Working Tips from Filmmakers Using Replicate
- 01 Read the price on each model page before a batch: some video models bill per second of output (Wan 2.1 i2v at $0.09/s for 480p, $0.25/s for 720p) and some images bill per output
- 02 Install the hosted MCP server in Claude or Cursor so you can compare and run models in plain language, instead of writing request code for each test
- 03 Test a model with 2 or 3 short generations before you queue a large batch; cold starts and per-second billing make big runs expensive to redo
- 04 For a deadline, keep a dedicated deployment warm for that model; you pay for idle time, so switch it off afterwards
Replicate pricing in 2026
| Plan | Price | What you get |
|---|---|---|
| Public models, hardware time | $0.000025/s (CPU) to $0.001525/s (H100) per second of runtime |
|
| Public models, per output | varies by model, e.g. flux-1.1-pro $0.04 per image; wan-2.1 i2v 480p $0.09/s; 720p $0.25/s per output image, token or second of video |
|
| Private / dedicated deployments | not listed on the pricing page setup + idle + active time (fast-booting fine-tunes: active time only) |
|
| Enterprise / committed spend | custom contract |
|
Prices checked against Replicate's pricing page on 9 Oct 2026. Vendors change plans often, so confirm before you buy.
Replicate vs ComfyUI
| Replicate | ComfyUI | |
|---|---|---|
| Price | Pay per second of hardware: $0.000225/s (T4) to $0.001525/s (H100). No monthly plan on the pricing page; enterprise is custom. | Local app free (GPL-3.0); Comfy Cloud Standard $20/mo ($16/mo billed yearly) |
| Infinite canvas | No | Partial |
| Node / flow builder | Not confirmed | Yes |
| AI agent | No | Partial |
| Chat assistant | No | Yes |
| Team collaboration | Not confirmed | Partial |
| Text to image | Yes | Yes |
| Image editing | Not confirmed | Yes |
| Text to video | Yes | Yes |
| Image to video | Yes | Yes |
| Video to video | Not confirmed | Partial |
| Native audio in video | Yes | Not confirmed |
| Voice / TTS | Yes | Not confirmed |
| Music & SFX | Yes | Yes |
| Upscaling | Not confirmed | Partial |
| Custom training | Yes | Partial |
| Third-party models | Yes | Yes |
| Public API | Yes | Yes |
| MCP server | Yes | Yes |
| Desktop / local | Partial | Yes |
| Open source | Partial | Yes |
| Free tier | Not confirmed | Yes |
| Commercial use | Not confirmed | Partial |
| Pricing model | Per second of GPU time, or per output for some models | Monthly subscription credits; desktop app free |
| Entry paid price | $0.000225/s on T4 (about $0.81/hr) | Standard $16/mo billed yearly |
| Top GPU listed | H100 at $0.001525/s; up to 8x H100 | RTX 6000 Pro, 96GB VRAM (Comfy Cloud) |
| Public API | Yes: REST + SDKs | Yes (Comfy API) |
| Visual node graph | No visual builder on site or docs | Yes: node graph is the core interface |
| Runs locally | Only self-hosted via Cog | Yes, desktop app is free |
Each column comes from that tool's own researched page. "Not confirmed" means the vendor doesn't say.
Open the full Replicate vs ComfyUI comparison →Use Cases
Integrations
Tags
Sources
Facts on this page were checked against Replicate's own pricing, product and changelog pages on , then cross-checked with independent coverage where vendor pages were silent. Anything we couldn't confirm is marked "Not confirmed" rather than guessed.
- Replicate pricing page replicate.com
- Replicate changelog replicate.com
- Replicate home replicate.com
- Replicate MCP server docs replicate.com
- Seedance 2.5 model page on Replicate replicate.com
- Cloudflare blog: Replicate joins Cloudflare blog.cloudflare.com
- Comfy Cloud pricing (comparison) comfy.org
Spotted something out of date? Tell us and we'll re-check it.
Discussion
No comments yet — be the first.