Replicate Review 2026: Pricing, Pros, Cons and Alternatives
How Replicate's two meters work (hardware seconds and per-output prices), three jobs priced on each, and what a dedicated GPU costs there next to Modal, RunPod, fal and Hugging Face.
Replicate is worth it in 2026 when you call public models in bursts and do not want to run GPUs. Most models bill per second of hardware (an L40S is $0.000975 a second) and some bill per output: FLUX 1.1 Pro is $0.04 an image, so 1,000 images cost $40, read on Replicate's pricing page on November 4, 2026.
The cost turns when you need your own always-on model. A deployment bills idle time too, so one L40S kept online for 30 days costs $2,527.20 by our arithmetic. The same GPU is $784.80 for the month on a RunPod Secure Cloud pod and $1,296 on a Hugging Face Inference Endpoint.
Key factsBest for: Developers calling public image, video and language models…
- Model catalog and runtime that bills public models by the second or by the output, with free idle time.
- Billing: Prepaid credit, spent per second of hardware or per output, depending on the model.
- Public models: Setup and idle time are free; you pay only while your request runs.
- Deployments: Setup, idle and active time are all billed, like a rented GPU.
- Ownership: Part of Cloudflare since November 2025, with the API unchanged.
Pros
- Public models bill only active seconds, so a quiet week costs almost nothing.
- Per-output models such as FLUX 1.1 Pro give an exact budget: 25 images per dollar.
- Every model page shows its hardware and an approximate cost per run before you call it.
Cons
- A dedicated L40S costs $2,527.20 for a 30-day month, more than three times a RunPod pod.
- At zero balance new work stops and your instances shut down unless auto reload is on.
- Shared public queues can cold-boot when other customers load the same model.

Replicate is the model catalog developers use when they want an image, video or language model behind one API call and no GPU to babysit. Its pricing has two meters, and which one a model uses decides whether your bill tracks compute seconds or finished outputs. This review explains both with priced examples, then shows where a dedicated GPU elsewhere is the cheaper home.
Toolradar data: on Toolradar's GPU cloud ranking, which tracks 44 tools as of November 3, 2026, Replicate is the most cross-shopped GPU cloud tool, listed as an alternative to 40 others, and the most compared one, with 261 side-by-side comparisons in 90 days. Buyers weigh it against everything, so the useful question is which of its meters fits your traffic.
All Replicate rates below come from its pricing page, docs and model pages; rival rates come from their own pricing pages, all read November 4, 2026. Every total is our arithmetic on list prices.
What is Replicate?
Replicate is a hosted platform for running machine learning models through an API. The catalog holds thousands of open-source models contributed by the community plus proprietary ones such as FLUX 1.1 Pro, and you can package your own with Cog, its open-source packaging tool, and run it on Replicate's GPUs next to the public catalog.
Replicate announced on November 17, 2025 that it was joining Cloudflare. The post says Replicate stays a distinct brand, the API is not changing, and existing models keep running, with tighter ties to Cloudflare's developer platform planned. If your app already lives on Workers, our Cloudflare Workers pricing page covers that side of the bill.
It is not a raw GPU rental: you buy compute by the second through models, so for steady self-hosted serving the honest comparison is a pod or endpoint host, and this review prices both.
How Replicate works
A public model runs on a shared hardware pool. According to Replicate's billing docs, you pay only while the model is processing your request; setup and idle time are free. The trade-off is the shared queue, so you can hit cold boots or scaling limits when other customers are busy on the same model.
A private model or a deployment gets its own instances and queue, and you pay for every second they are online: setup, idle and active. Replicate scales them down to a minimum you set (zero by default) and keeps an instance idle for a few minutes between bursts. Fast booting fine-tunes are the exception, billed only while active even when private.
Billing is prepaid for most accounts. You buy credit, usage is deducted, and the prepaid credit page says purchased credit is valid for one year and is not refundable. When the balance hits zero, Replicate stops new work and shuts down the infrastructure it runs for you, so turn on auto reload (minimum threshold $5, minimum reload balance $15) before a production launch. Failed runs on public models are free; a canceled run on an official model can still be charged.
Replicate key features
Replicate pricing
Replicate publishes its hardware rates on its pricing page, read November 4, 2026: a T4 is $0.000225 a second, an L40S $0.000975, an A100 80 GB $0.0014 and an H100 $0.001525, with multi-GPU A100, H100 and H200 sizes up to 8 GPUs (larger counts and H200 capacity need committed spend contracts). Volume discounts exist for large spend, with no published rate.
Example 1, per output. FLUX 1.1 Pro is priced per image on its model page, 25 images per dollar. A batch of 1,000 product shots costs $40 regardless of how long each one takes. Video works the same way: Wan 2.1 image-to-video at 720p is $0.25 per second of output, so a 5-second clip is $1.25.
Example 2, per second on a public model. SDXL runs on an L40S and typically completes within 4 seconds; Replicate estimates about $0.0038 a run. A batch of 1,000 images is about $3.80, roughly a tenth of the FLUX batch, because you pay seconds, not a per-image fee. The catch is that a slow input or a cold boot raises the cost, while the per-output price stays fixed.
Example 3, a dedicated deployment. Keep one L40S online for a 30-day month and the meter runs 720 hours at $3.51: $2,527.20, busy or not. An H100 for the same month is $3,952.80. Below about 924 SDXL-sized runs an hour, the shared public model is cheaper than holding that L40S, by our arithmetic on Replicate's own estimate.
Elsewhere, the same 720 hours on an L40S cost $784.80 on a RunPod Secure Cloud pod, $1,260 on RunPod Serverless at its 48 GB rate, $1,296 on a Hugging Face Inference Endpoint and $1,404.86 on Modal before its monthly credit and separate CPU and memory charges. For FLUX 1.1 Pro, fal charges the same amount per megapixel, rounded up, so a 1 MP image costs what it does on Replicate and a larger one costs more. Our cheapest GPU cloud guide ranks the hourly hosts in more depth.
| Plan | Price | Best for |
|---|---|---|
| CPU (Small) / CPU | $0.000025 / $0.0001 per sec | 1 or 4 vCPUs with 2 or 8 GB RAM, for light models and preprocessing |
| Nvidia T4 | $0.000225 per sec ($0.81/hr) | 16 GB GPU for small image and audio models |
| Nvidia L40S | $0.000975 per sec ($3.51/hr) | 48 GB GPU; the hardware behind SDXL and many diffusion models |
| Nvidia A100 80 GB | $0.0014 per sec ($5.04/hr) | 80 GB GPU, 10 vCPUs and 144 GB RAM; 2x, 4x and 8x sizes listed |
| Nvidia H100 | $0.001525 per sec ($5.49/hr) | 80 GB GPU for large models; multi-GPU H100 and H200 on contract |
| FLUX 1.1 Pro | $0.04 per image | Per-output pricing; 25 images per dollar |
| Wan 2.1 image-to-video | $0.09 (480p) / $0.25 (720p) per sec of video | Billed by seconds of finished footage, not GPU time |
| Enterprise | Custom | Volume discounts, higher GPU limits, performance SLAs, dedicated account manager |
| Modal (L40S) | $0.000542 per sec | Serverless containers; $30 monthly credit on Starter; CPU and memory metered on top |
| RunPod pod (L40S) | $1.09/hr Secure Cloud ($0.79 Community) | Dedicated GPU pod billed by the hour, no model catalog |
| Hugging Face Inference Endpoint (L40S) | $1.80/hr | Dedicated autoscaling endpoint on AWS for any Hub model |
| fal (H100 / FLUX 1.1 Pro) | $4.50/hr list; $0.04 per megapixel | Model APIs billed per output plus a GPU fleet for custom apps |
Replicate pros and cons
What we like
- Public models bill only active seconds, so a quiet week costs almost nothing.
- Per-output models such as FLUX 1.1 Pro give an exact budget: 25 images per dollar.
- Every model page shows its hardware and an approximate cost per run before you call it.
What could be better
- A dedicated L40S costs $2,527.20 for a 30-day month, more than three times a RunPod pod.
- At zero balance new work stops and your instances shut down unless auto reload is on.
- Shared public queues can cold-boot when other customers load the same model.
Who Replicate is for
Replicate fits developers who call models, not developers who host them. A prototype, a feature that generates images on demand, or a batch job that runs a few times a week all sit well on public models, where idle time is free and the model page tells you the price before you run it.
It also fits teams that want proprietary and open models behind one API key and one bill, since partner models such as FLUX 1.1 Pro sit next to community models in the same catalog. For language models, compare the per-token rates against direct providers in our LLM APIs ranking and OpenAI API pricing breakdown first.
Skip Replicate for a custom model that serves traffic all day. A deployment bills every online second, and an hourly pod or endpoint costs a fraction of that, as the comparison above shows. Skip it too for local experiments, where Ollama runs open models on your own machine for free.
Best Replicate alternatives
If Replicate is not the right fit, these are the closest options.
| Tool | Best for | Starts at | |
|---|---|---|---|
| Replicate | Developers calling public image, video and language models in bursts through one API. | Pay as you go from prepaid credit: L40S $0.000975/sec, H100 $0.001525/ | Visit → |
| Modal | Python teams deploying their own models with spiky traffic and a code-first workflow. | Free plan · Team $250/mo · custom-priced plans too | Visit → |
| RunPod | Teams serving their own model around the clock who want the lowest hourly GPU price. | L40S pod $1.09/hr Secure Cloud ($0.79 Community) | Visit → |
| fal | Apps built on image and video generation that want a large hosted catalog and fast inference. | Model APIs per output (FLUX 1.1 Pro $0.04/MP) | Visit → |
| Hugging Face Inference Endpoints | Teams deploying Hub models on dedicated hardware with a fixed hourly price. | Free plan · Pro $9/mo · Team $20/mo per user · custom-priced plans too | Visit → |
The bottom line
Replicate is worth it as a pay-per-call model API and expensive as a GPU landlord. Use public and per-output models for bursty work, where idle time is free and costs are known per run. Once your own model needs to stay online most of the day, move it to a RunPod pod, a Hugging Face endpoint or Modal, which charge less for the same hardware.
Read the Modal review for the serverless container route, the serverless platforms ranking for the wider field, cloud free tiers compared for the free allowances, free LLM APIs if you only need text, and the AI agent sandboxes ranking for agent runtimes. Toolradar keeps profiles of Replicate, RunPod, Modal and fal. Subscribe free to get the next GPU price change in the Devshot brief.
Methodology: Replicate's hardware rates, billing rules and model prices were read on its pricing page, docs and model pages on November 4, 2026, and rivals' rates on their pricing pages the same day. The three jobs are our assumptions, the 30-day month is 720 hours, and every total is our arithmetic on list rates before tax. Devshot does not run platforms hands-on.
Cite this: Devshot, "Replicate Review 2026: Pricing, Pros, Cons and Alternatives", November 2026.
Frequently asked questions
Is Replicate worth it in 2026?
How much does Replicate cost?
Is Replicate free?
Why is my Replicate deployment so expensive?
Replicate vs Modal: which is cheaper?
Some offers on this page may be paid placements or contain affiliate links.
Get the Devshot brief
Free daily newsletter, read in 5 minutes.
Subscribe free