Replicate pricing guide details
Pricing verified against replicate.com as of April 2026. Replicate bills per second of compute with no subscription required.
Pricing Overview
Replicate uses a pure pay-as-you-go pricing model billed per second of compute time. There are no monthly subscriptions, seat licenses, or minimum commitments. You pay only for the hardware seconds your model predictions consume. This usage-based approach makes Replicate accessible for experimentation while scaling costs linearly with production workloads.
Hardware pricing ranges from $0.09/hr for CPU instances to $43.92/hr for 8x H100 GPU clusters. Public models hosted on Replicate have fixed per-prediction pricing: Flux Schnell costs $0.003/image, Flux 1.1 Pro costs $0.04/image, and DeepSeek R1 runs at $3.75 per 1M input tokens. Video generation with Wan 2.1 at 480p costs $0.09 per second of generated video. Enterprise customers can negotiate volume discounts through committed spend agreements.
Plan Comparison
Replicate does not use traditional subscription tiers. Instead, pricing is determined by the hardware tier selected for each model deployment:
| Hardware Tier | Hourly Rate | Per-Second Rate | Best For |
|---|---|---|---|
| CPU | $0.09/hr | $0.000025/sec | Lightweight preprocessing, text models |
| Nvidia T4 | $0.81/hr | $0.000225/sec | Budget inference, small image models |
| Nvidia A40 Large | $1.48/hr | $0.000411/sec | Mid-range inference workloads |
| A100 40GB | $3.15/hr | $0.000875/sec | Large language models, training |
| A100 80GB | $5.04/hr | $0.001400/sec | 70B+ parameter models, high-memory tasks |
| H100 | $5.49/hr | $0.001525/sec | Fastest single-GPU inference |
| 4x H100 | $21.96/hr | $0.006100/sec | Distributed inference, large batch jobs |
| 8x H100 | $43.92/hr | $0.012200/sec | Maximum throughput, multi-GPU training |
Hidden Costs and Considerations
Private-model instance time. Most private models run on dedicated hardware. Replicate bills for all time an instance is online: setup, idle time while waiting for requests, and active request-processing time. This means costs can continue outside active inference work.
Fast-booting fine-tunes are different. Fast-booting fine-tunes are billed only while active and processing requests, so idle time is not billed. Replicate identifies these models as fast-booting fine-tunes in the model version list.
Public-model billing varies. Most public models are billed for the time they take to run, with the per-second rate determined by the hardware used. Some public models are instead billed by input and output; Replicate provides estimates on each model’s page.
Capacity and enterprise terms. Some additional multi-GPU hardware capacity is available with committed-spend contracts. Replicate also states that volume discounts are available for large amounts of spend, while dedicated account management, priority support, higher GPU limits, performance SLAs, and onboarding or optimization help are offered for enterprise and complex requirements. The supplied evidence does not disclose discount levels, committed-spend amounts, or the terms required for those options.
Cost Estimates by Team Size
Solo developer or hobbyist: Running 1,000 image generations per month with Flux Schnell at $0.003/image costs $3/month. Occasional experimentation with larger models on T4 GPUs ($0.81/hr) for 10 hours adds $8.10. Monthly total: $11-$15.
Small startup (3-5 engineers): A team running 50,000 Flux Schnell predictions per month ($150) plus a custom model on A100 80GB for 100 hours ($504) with DeepSeek R1 processing 10M tokens ($37.50). Monthly total: $500-$900.
Mid-size company (15-25 engineers): Production workloads running custom models on H100 GPUs for 500 hours/month ($2,745), plus 500,000 image generations ($1,500), and video generation processing 1,000 seconds ($90). Before enterprise discounts, monthly total: $4,000-$6,000. With committed spend discounts, expect 15-25% savings.
How Replicate Pricing Compares
Replicate's per-second billing model differs fundamentally from competitors that charge per token or per million tokens. Direct cost comparison depends on the specific model and workload pattern.
vs. Fireworks AI: Fireworks charges per token, starting at $0.10 per 1M tokens for sub-4B parameter models and $0.90 per 1M tokens for 16B+ models. For LLM inference, Fireworks is substantially cheaper for high-throughput text workloads. Replicate's advantage is broader model support including image, video, and audio models that Fireworks does not host.
vs. Together AI: Together AI offers inference from $0.10 per 1M tokens for smaller models. For pure LLM serving, Together provides more predictable per-token pricing. Replicate's per-second hardware billing can be more cost-effective for models with variable output lengths or non-text modalities.
vs. Groq: Groq charges $0.59 per 1M input tokens and $0.79 per 1M output tokens for Llama 70B. For LLM-only workloads requiring low latency, Groq undercuts Replicate on price and speed. Replicate serves a broader set of use cases beyond text generation.
Replicate's strongest cost advantage is for teams that need to run custom models (fine-tuned or proprietary) across multiple modalities. The per-second billing model works well for bursty, unpredictable workloads where you want to avoid paying for idle capacity. For teams focused purely on LLM inference at scale, token-based providers like Fireworks AI and Together AI deliver better unit economics.