Replicate Pricing in 2026: Plans, Limits and Real Costs
Start with a public model charged per output or processing time. Most founders should begin with public models; choose a private model only when dedicated capacity or custom deployment is required, and expect to pay for setup and idle time.
Verified 2026-10-27 · official pricing page · Replicate in the directory
Quick pricing verdict
- Starting price
- There is no subscription or flat starting price. Public models are charged by output or processing time, while private models are charged for hardware runtime.
- Free plan
- Not published
- Best plan for most people
- Public models are the default choice because usage is billed based on how the model is consumed.
- Plan to avoid first
- Avoid dedicated private models until you need custom deployment or dedicated hardware, because most charge for setup, idle, and active time.
- Biggest hidden cost
- Most private models remain billable while idle, including setup time and time spent waiting for requests.
- Pricing checked
- 2026-10-27, against the official page
Replicate has no published subscription tiers. Public models use usage-based pricing, generally by output unit or processing time, and model pages provide cost estimates.
Private models run on dedicated hardware and usually charge for the full time instances are online: setup, idle, and active processing. Fast booting fine-tunes are billed only while active.
Plans compared
| Plan | Price | Best for | Key limits |
|---|---|---|---|
| Public models | $0.01 / thousand output tokens; $3.75 / million input tokens and other per-output examples | Trying and running public models without deploying custom infrastructure | Pricing varies by model; most are billed by run time, while some are billed by input and output. |
| Private models | $0.000025/sec to $0.012200/sec depending on hardware | Deploying custom models with Cog and obtaining dedicated hardware | Most private models bill setup, idle, and active time. Fast booting fine-tunes bill only active processing time. |
| Enterprise & volume discounts | Not published | Large spend, complex requirements, and teams needing additional support | Committed spend contracts are required for some additional Multi-GPU A100, Multi-GPU H100, and H200 capacity. |
Source: Replicate pricing page, checked 2026-10-27. See the archived page we read.
What each plan includes
Public models
- Includes
- Thousands of open-source machine learning models contributed by the community, plus a variety of proprietary models. Models may be billed by processing time, output unit, or input and output.
- Missing
- Free allowance, monthly cap, and universal included usage are not published.
- Avoid if
- You need a custom model on dedicated hardware or do not want usage-based charges.
- Verdict
- Best starting point for most solo founders.
Private models
- Includes
- Custom model deployment using Cog, dedicated hardware for most models, automatic scaling for traffic, and active-time-only billing for models labeled as fast booting fine-tunes.
- Missing
- A fixed monthly private-model fee, minimum spend, and universal idle-time limits are not published.
- Avoid if
- Traffic is intermittent and you cannot justify paying while the model is idle.
- Verdict
- Choose when custom deployment or dedicated hardware is worth the runtime cost.
Enterprise & volume discounts
- Includes
- Dedicated account manager, priority support, higher GPU limits, performance SLAs, onboarding help, custom model help, optimizations, and volume discounts for large spend.
- Missing
- Exact account fees, discount rates, and minimum spend are not published.
- Avoid if
- You are an individual founder with limited usage and do not need enterprise support.
- Verdict
- Relevant after usage or support requirements justify a negotiated arrangement.
Add-ons, overages, and surprise costs
| Cost item | Trigger | Published amount | Billing behaviour |
|---|---|---|---|
| Private-model idle time | The model instance is online but waiting for requests | Charged according to the selected hardware price | Most private models remain billable during idle time. |
| Private-model setup time | The model instance is setting up | Charged according to the selected hardware price | Most private models bill setup time along with idle and active processing time. |
| Private-model active processing time | The model is processing requests | $0.000025/sec to $0.012200/sec depending on hardware | You pay for the duration of each online model instance. |
| Committed-spend capacity | Using additional Multi-GPU A100, Multi-GPU H100, or H200 capacity | Not published | The listed capacity is available with committed spend contracts. |
| Variable model charges | Running public models | $0.01 / thousand output tokens to $3.75 / million input tokens and other per-output examples shown | Each model has its own billing unit, so repeated requests can produce continuing charges. |
Source: Platform pricing docs, checked 2026-10-27. See the archived page we read.
Which plan should you choose?
- Solo founder testing images
- black-forest-labs/flux-schnell — It is listed at $3.00 / thousand output images.
- Solo founder prioritizing image quality
- black-forest-labs/flux-1.1-pro — It is listed at $0.04 / output image.
- Solo founder using a reasoning model
- deepseek-ai/deepseek-r1 — It is listed at $0.01 / thousand output tokens and $3.75 / million input tokens.
- Solo founder creating image-to-video output
- wavespeedai/wan-2.1-i2v-480p or wavespeedai/wan-2.1-i2v-720p — The published rates are $0.09 / second of output video and $0.25 / second of output video, respectively.
- Founder deploying a custom model
- Private models with a fast booting fine-tune when available — Fast booting fine-tunes are billed only for active processing time, without idle-time charges.
Frequently asked questions
- Does Replicate publish monthly subscription plans?
- No. The page states that you only pay for what you use.
- How are public models billed?
- Most are billed by processing time. Some are billed by input and output, with pricing varying by model.
- Are private models billed while idle?
- Yes for most private models. They bill for setup, idle, and active processing time. Fast booting fine-tunes are billed only while active.
- Can I estimate the cost of a public model?
- Yes. The page says each model page includes estimates of how much the model will cost.
- What enterprise benefits are stated?
- Dedicated account manager, priority support, higher GPU limits, performance SLAs, onboarding help, custom model help, optimizations, and volume discounts for large spend.
- What hardware pricing is listed?
- CPU (Small) is $0.000025/sec, CPU is $0.000100/sec, Nvidia A100 (80GB) is $0.001400/sec, Nvidia H100 is $0.001525/sec, Nvidia L40S is $0.000975/sec, and Nvidia T4 is $0.000225/sec.
What we read
Every figure on this page was extracted from the vendor's own pricing page at the moment below. The captured text is one click away, so you can check our reading against what they publish today.
Source captured 2026-10-27 16:17 UTC · read the captured text
Pricing change log
What moved since we started tracking this page, newest first. A price that changes silently is the failure mode of every other pricing guide.
- 2026-10-27First published, extracted from https://replicate.com/pricing; archive snapshot 2026-10-25.