Replicate API Pricing Calculator 2026
Estimate your monthly AI model inference costs for public open-source models, private deployments, and official flat-rate APIs on Replicate’s cloud platform.
Inference Budget Planner
Define your model requirements and runtime volume
Select the computer setup utilized by your target machine learning model.
Average volume of inference requests executed per day.
Estimated processing seconds per prompt (setup/idle is free on public models).
Your Cost Breakdown
Estimated Replicate bill predictions
Select your deployment tier, configure active hardware volume, then click Calculate API Cost to construct a tailored 2026 expense profile.
Replicate API Hardware pricing
2026 baseline hardware pricing tiers and specialized output models.
| Hardware / Model | On-Demand Rate | Notes |
|---|---|---|
| CPU | $0.000100 / sec ($0.36/hr) | Best for testing or lightweight non-GPU tasks |
| Nvidia T4 GPU | $0.000225 / sec ($0.81/hr) | Affordable tier for lighter model testing |
| Nvidia L40S GPU | $0.000975 / sec ($3.51/hr) | Highly efficient modern general inference GPU |
| Nvidia A100 (80GB) | $0.001400 / sec ($5.04/hr) | Sustained large-scale model workflows & fine-tuning |
| Nvidia H100 GPU | $0.001525 / sec ($5.49/hr) | Maximum throughput with next-generation speeds |
| FLUX.1 [schnell] | $0.003 / image | Flat rate, zero billing overhead for cold boots |
| FLUX.1 [dev] | $0.025 / image | Output flat rate for creative styling outputs |
| FLUX 1.1 Pro | $0.040 / image | Production-grade professional generations |
| Recraft V3 | $0.040 / image | Vector graphic designs & stylized image tasks |
| Wan 2.1 Video (480p) | $0.080 / output second | Highly accelerated video generation pipeline |
Replicate Pricing FAQ
Everything developers and startups need to know about calculating and optimizing active Replicate spend.
Replicate operates on a strictly pay-as-you-go billing foundation. For public catalog models, you only pay for the exact active run-time seconds execution. Private custom models and deployments bill for all online runtime—meaning you also pay for setup times and idle holding windows to secure instant warm-starts.
Public models run on a shared cluster, billing you only for active processing time. However, they are prone to “cold start” latency spikes if they haven’t been run recently. Private deployments run on dedicated instances that stay online continuously, eliminating cold starts at the cost of billing for all active and idle instance uptime.
Launched to provide more predictable budgets, Official Models (such as FLUX image generators) bypass complex GPU per-second compute calculations. Instead, they feature flat-rate, output-based pricing (e.g., $0.003 per image for FLUX.1 [schnell]). This removes setup time overhead costs entirely.
For standard public and official models, Replicate does not bill you for execution requests that fail. However, on private models and custom deployments, you are billed directly for hardware runtime, which means failed script executions still count towards your active online instance runtime billing.
Replicate is highly cost-efficient for testing, small-scale deployments, or variable traffic spikes where you scale to zero and pay nothing for idle time. However, for high-volume, 24/7 sustained production pipelines, a dedicated bare-metal GPU server with fixed monthly pricing often provides significant savings compared to paying by the second.