Replicate
Replicate runs and fine-tunes open-source AI models via API. See 2026 per-second GPU pricing, the Cloudflare acquisition, and real alternatives.
About Replicate
What is Replicate (2026 update)
Replicate is a cloud API platform in the art and creative design category that lets developers run, fine-tune, and deploy open-source machine learning models without managing their own GPU infrastructure. It hosts more than 50,000 production-ready models that can be called with a few lines of code instead of setting up a training and inference pipeline from scratch.
What is new in 2026
The biggest 2026 development for Replicate is that Cloudflare announced on November 17, 2025 that it had agreed to acquire the company, with the deal closing December 1, 2025. The acquisition folds Replicate’s model catalog into Cloudflare Workers AI, so developers building on Cloudflare can access Replicate’s models through Cloudflare’s platform, though as of mid-2026 Replicate has not introduced new Cloudflare-specific pricing tiers. On the product side, 2026 additions include Deployments, which are dedicated GPU instances for production workloads, fine-tuning workflows for models like SDXL and Llama, and a Python SDK that integrates with popular ML frameworks. Cold start times have also improved significantly, with most models now responding in under 5 seconds after the first call, and Cog 0.14.0 added support for models that run predictions concurrently via async and await.
Key features
- Access to more than 50,000 open-source and community models through one API
- Deployments for dedicated GPU instances that keep production workloads warm
- Fine-tuning workflows for models such as SDXL and Llama
- Python SDK that integrates directly with common ML frameworks
- Organizations for sharing models, API tokens, billing, and dashboards across a team, now creatable without a linked GitHub organization
- Cog, Replicate’s open-source tool for packaging models to run in containers
Pricing in 2026
Replicate has no flat subscription tiers. It bills per second of GPU or CPU time for most public models, while some popular models like FLUX and DeepSeek R1 use fixed per-output pricing instead.
| Plan | Price | What you get |
|---|---|---|
| CPU (small) | $0.000025/sec ($0.09/hour) | Lightweight CPU inference for simple models |
| CPU (standard) | $0.000100/sec ($0.36/hour) | General CPU inference |
| Nvidia T4 GPU | $0.000225/sec ($0.81/hour) | Entry-level GPU inference |
| Nvidia L40S GPU | $0.000975/sec ($3.51/hour) | Mid-tier GPU for larger models |
| Nvidia A100 (80GB) | $0.001400/sec ($5.04/hour) | High-memory GPU for large models and fine-tuning |
| Nvidia H100 GPU | $0.001525/sec ($5.49/hour) | Top-tier GPU for the most demanding models |
Multi-GPU configurations are billed proportionally higher, for example 2x A100 at $0.002800/sec or 4x H100 at $0.006100/sec. New accounts use prepaid credits valid for one year, and you are only billed for active processing time, since cold starts are not charged. Organizations needing higher usage limits, a dedicated account manager, or priority support can request enterprise and volume pricing directly from Replicate.
Who should use it
Replicate fits developers and small teams who want to call open-source image, video, audio, or language models through an API without buying or managing GPUs themselves. It suits both experimentation, since you only pay for what you use, and production workloads, since Deployments keep dedicated instances warm for consistent latency. It is a weaker fit for teams that need a fixed monthly budget rather than variable usage-based billing, since costs scale directly with GPU time consumed.
Limitations and alternatives
Because Replicate bills by GPU time, costs can be harder to predict than a flat subscription, especially for models billed by the second rather than by fixed output pricing, so teams running heavy workloads should model expected usage before committing. The Cloudflare acquisition also introduces some uncertainty about how pricing and product direction might change as integration with Cloudflare Workers AI deepens. For running a single open-source segmentation or vision model without the broader Replicate catalog, compare against a narrower hosted option like Hugging Face Inference Endpoints, and for teams that want fully managed GPU infrastructure with more control, RunPod and Modal are common alternatives developers evaluate alongside Replicate. Check each provider’s current per-second or per-hour GPU rates directly, since compute pricing changes with hardware availability.
Frequently asked questions
How does Replicate pricing work?
Replicate bills per second of GPU or CPU time you use, with rates ranging from $0.000025/sec for small CPU tasks to $0.001525/sec for an Nvidia H100 GPU. Some models like FLUX use fixed per-output pricing instead, and cold starts are not charged.
Did Cloudflare acquire Replicate?
Yes, Cloudflare announced its agreement to acquire Replicate on November 17, 2025, with the deal closing December 1, 2025. As of mid-2026 no new Cloudflare-specific pricing tiers had been introduced.
Does Replicate have subscription plans?
No, Replicate uses usage-based billing rather than flat subscription tiers, though organizations with higher usage needs can request custom enterprise and volume pricing.
Can I fine-tune models on Replicate?
Yes, Replicate offers fine-tuning workflows for models such as SDXL and Llama, alongside Deployments for keeping dedicated GPU instances warm for production use.
Information last verified: September 2026.
Related Tools
