Replicate lets you run thousands of open-source AI models — image, video, audio and language — with one API call, billing by the second of GPU time. You don't manage servers; you pay only while a model runs. Here's what it costs and how to get your token.
| Hardware | ≈ $/second | ≈ $/1,000 runs* | Best for |
|---|---|---|---|
| CPU | $0.0001 | ~$5 | Light pre/post-processing |
| Nvidia T4 GPU | $0.000225 | ~$25 | Small image / audio models |
| Nvidia A100 (40GB) | $0.00115 | ~$140 | LLMs, SDXL, video |
| Nvidia H100 | $0.00250 | ~$300 | Large / fast inference |
→ Running language models? Compare token-priced APIs on the AI API cost calculator — sometimes a hosted LLM is cheaper than per-second GPU.
Replicate gives a small amount of free credit to try models, then it's pay-as-you-go by the second with no monthly minimum. There's no perpetual free tier, but you only pay while a model actually runs, so idle costs are zero.
1. Sign up at replicate.com (GitHub login works).
2. Open Account → API tokens and copy your token.
3. Add a payment method for usage beyond the free credit.
4. Call any model by its version id — billing starts when the run starts, stops when it ends.
Test it with a simple request:
For hosted LLMs by token (often cheaper than per-second GPU), see OpenAI, DeepSeek or Together AI. For image generation specifically, Stability AI and Fal.ai compete on price; for raw model hosting, Hugging Face Inference Endpoints is the closest rival.
You pay per second of compute while a model runs, at a rate set by the hardware (CPU, T4, A100, H100). No monthly fee; idle time costs nothing.
For language tasks, a token-priced API like DeepSeek or GPT-4o mini is usually cheaper and simpler. Replicate wins when you need a specific open model (image, video, audio) that hosted APIs don't offer.
A small free trial credit, then pay-as-you-go. No perpetual free tier.
Sign up at replicate.com, open Account → API tokens, and copy your token.
Not affiliated with Replicate. Prices are reference estimates — always verify on the official pricing page.