Replicate

💰 Pricing & Plans

🔧 API & Developer Experience

• •Prediction API: REST endpoint at https://api.replicate.com/v1/predictions — POST to create an async prediction, GET /predictions/{id} to poll, or stream via the official Python/Node clients. Every model on Replicate is callable through this one canonical interface.
• •Model versions & pinning: Each model exposes immutable version hashes (e.g. stability-ai/sdxl:39ed52b2...). You can pin to a specific version for reproducibility or float to the latest — production deployments should always pin.
• •Webhooks: Register a webhook URL per prediction; Replicate POSTs the result back when the run completes, so long-running video / 3D jobs don't need to be polled.
• •Streaming outputs: Diffusion and LLM models can stream intermediate frames / tokens over SSE through the official clients. Useful for image previews before a job finishes or for token-by-token LLM output.
• •Bring your own Cog container: Replicate's open-source Cog tool lets you package any custom model as a Docker container and deploy it onto Replicate hardware with one command — no infra work needed to onboard a new model.
• •Auth & SDKs: API token issued from replicate.com/account; official Python (replicate) and Node (@replicate/replicate) clients plus a thin REST API. No first-party Go / Java SDK; community wrappers exist.
• •Rate limits: Soft-throttled by per-account concurrency (default ~10 concurrent predictions, raised with usage history). Burst limits apply to /v1/predictions POSTs — long video jobs count against the concurrency pool, not against a request-per-minute budget.

🎬 50,000+ Community Models on Per-Second GPU Billing

Replicate's defining capability is the combination of a per-second GPU-time billing model with a 50,000+ community-curated model catalog: instead of token-based pricing, every prediction is metered in seconds of GPU runtime on the underlying hardware (A100 40GB / 80GB, H100, T4, etc.), and the model itself is a thin Cog container that boots on demand. This makes the platform genuinely hardware-flexible — you can run a 7B LLM, a Stable Diffusion XL image generator, a Whisper transcription, a MusicGen audio model, or a custom diffusion checkpoint through the same /v1/predictions endpoint, with the runtime chosen per model based on its compute needs. The trade-off is cost predictability: a fast model on cheap hardware is dramatically cheaper than a slow model on H100s, so per-prediction bills vary by orders of magnitude depending on which model you call. For prototyping and experimentation across modalities — image, video, audio, 3D, embeddings — this breadth and on-demand GPU model is unmatched; for steady high-throughput LLM traffic, a token-priced provider is usually more economical.

🌐 China Access & Latency

Replicate's API (api.replicate.com) and dashboard are not directly reachable from mainland China: the domain is blocked at the network edge and the platform does not publish a China-optimized mirror or any official CN endpoint. Access from China requires either a stable overseas proxy / VPN, or routing through a third-party aggregator that re-sells Replicate access (such as SiliconFlow or some OpenRouter routes for selected models). Latency from China on a good proxy typically sits in the 300-700 ms RTT range before any inference starts, and individual predictions add seconds-to-minutes of GPU runtime on top depending on the model class — image generation runs 2-10s, video diffusion runs 30s-3min, 3D mesh generation can run several minutes. For production workloads serving China-based users, Replicate is generally a low-priority option; self-hosted open-source substitutes (Stable Diffusion WebUI, ComfyUI on domestic GPU clouds, or domestic multimodal APIs) usually win on both latency and cost. Most China-region teams that do use Replicate do so for prototyping or model evaluation, not for production traffic.