Replicate

Run open-source AI models without managing infrastructure

3.6(0 reviews)usage-basedFounded 2019United States

free tiergpu availableus datacenterapi access

Integrates with

Replicate lets you run open source AI models via API. Thousands of community-contributed models. Pay per prediction. Run Stable Diffusion, Llama, Whisper, and more with one API call.

What Replicate actually is

Replicate is a hosted inference platform: you call an HTTP endpoint with an input, a machine-learning model somewhere in Replicate's fleet produces an output, and you are billed for the compute the run consumed. The catalog is the product. Thousands of community-contributed models — image generation, video generation, upscaling, speech, music, and text — are callable with one request, alongside proprietary models from vendors like Google and DeepSeek. If you want your own model, you package it with Cog, Replicate's open-source container format, push it, and it gets the same API treatment as everything else.

The company was founded in 2019 by Ben Firshman (previously the product lead for Docker Compose) and Andreas Jansson (ex-Spotify ML), launched publicly in 2023 with $17.8 million in funding, and raised a $40 million Series B led by Andreessen Horowitz later that year. In November 2025, Cloudflare announced an agreement to acquire Replicate, with the stated goal of folding its model catalog into Workers AI. The service continues to run under the Replicate brand; what the integration means for pricing and hardware over the next two years is an open question, and anyone building long-term on the platform should weigh that uncertainty honestly.

We place Replicate in /category/gpu-cloud and /category/serverless-platforms because it blurs them: the billing unit is GPU time, but the delivery model is serverless. You never provision a machine. Public models scale to zero when nobody is calling them, and idle time on public models costs nothing. For a prototype, a side project, or a product with spiky traffic, that is the correct economic shape.

Pricing: what you really pay

Two billing schemes coexist. Many public models are billed by the time they take to run, at a per-second rate set by the underlying hardware. The published hardware table runs: CPU small at $0.000025 per second ($0.09 per hour), CPU at $0.0001 per second ($0.36 per hour), Nvidia T4 at $0.000225 per second ($0.81 per hour), Nvidia L40S at $3.51 per hour, Nvidia A100 (80 GB) at $5.04 per hour, and Nvidia H100 at $5.49 per hour. Multi-GPU configurations scale linearly — 8x H100 lists at $43.92 per hour — with the largest capacities gated behind committed-spend contracts. Some models instead bill per unit of output: text models charge per thousand tokens (one reasoning model lists at $0.01 per thousand output tokens and $3.75 per million input tokens), video models charge per second of generated video (Google's Veo 3.1 lists at $0.40 per second with audio, $0.20 without), and some image models charge per thousand output images.

Three billing rules matter more than the headline rates. First, for public models you pay only while a prediction is running — setup and idle time are free, and a model scaled to zero costs nothing between requests. Second, failed runs are not charged. Third, since July 16, 2025, new accounts are billed through prepaid credit: you top up first, usage draws down the balance, purchased credit is valid for one year and is not refundable, and when the balance hits zero new work stops. Optional auto-reload triggers at a $5 minimum threshold with a $15 minimum reload.

Is that competitive? Against raw GPU rental, the sticker rates are mid-market: H100 at $5.49 per hour undercuts some inference specialists (Baseten lists $6.50 per hour for the same class) but sits above others (fal lists $4.50 per hour). The decisive variable is utilization. For bursty workloads, per-second billing with scale-to-zero wins decisively — you pay for seconds of work and nothing else. For steady, high-throughput serving, per-second billing accumulates faster than a warm dedicated instance would, and we would look at reserved capacity or a different provider before committing a 24/7 production load to it.

The free story is narrower than the DB tags suggest. Replicate lets new accounts run a curated set of "Try for Free" models a limited number of times without a card; the exact allowance is not published, and it is a one-time budget, not a monthly reset. Once it is gone, billing setup is required, and un-billed accounts are rate-limited to one request per second, six per minute. We treat this as a trial allowance rather than a free tier in the recurring sense.

Performance & infrastructure

Replicate does not sell regions; it routes requests to available GPU capacity automatically, and the API does not expose which region a given run executed in. Third-party assessments describe the infrastructure as US-only for data-residency purposes, with no EU data-center option and no customer-selectable region pinning. For GDPR-constrained workloads that is a real limitation, and we would not send European personal data through it without proper contractual controls.

Hardware coverage is genuinely broad for a serverless offering: T4 and L40S for lighter jobs, A100 (40 GB and 80 GB) and H100 for heavy training and inference, H200 behind committed-spend agreements, and multi-GPU configurations up to 8x H100. That range is why the catalog can include everything from a tiny upscaler to a large video diffusion model.

Operational behavior is predictable in one way and unpredictable in another. Predictable: per-second metering means a run costs what its runtime says it costs, model pages show typical run times and costs, and versioning lets you pin a specific model version so a community update cannot silently change your outputs. Unpredictable: cold starts. A large model loading onto a GPU can take minutes on a cold deployment, which is fine for async batch work and unacceptable for user-facing latency budgets. Webhooks and the queue-based API exist precisely because many runs are long; blocking HTTP calls are the wrong shape for video generation.

Data handling deserves a paragraph because it surprises people. By Replicate's own documentation, inputs, outputs, and logs for API-created predictions are automatically removed after an hour by default, while data for predictions created through the web interface is kept indefinitely. If your use case involves sensitive inputs, that default asymmetry is worth configuring deliberately. There are no published SOC 2 Type II or ISO 27001 certifications as of our last check, which matters more now that regulated enterprises are the natural acquirer's customer base.

Where it falls short

Cost at volume is the headline complaint, and it is legitimate. Per-second GPU billing is excellent economics at 5% utilization and mediocre economics at 90%. Teams that grow into steady traffic routinely discover they are paying a premium over renting the same hardware, and the migration path off Replicate means repackaging models and rewiring clients.

Then there is the platform-control list. No region selection, no dedicated capacity without a commercial agreement, no IPv6 story we could document, no published uptime SLA, and — despite the Cloudflare acquisition — no published enterprise-grade compliance certifications yet. The prepaid credit system has sharp edges: credits expire after a year, are non-refundable, and a zero balance halts new runs. A failed auto-reload on a Friday night is a self-inflicted outage.

The community catalog cuts both ways. Anyone can publish, so quality varies, licenses vary, and maintenance varies; a model you depend on can be updated, deprecated, or left stale by its publisher. Pinning versions mitigates this but does not remove the risk that the underlying container is no longer maintained. And while Cog-based custom models are a real strength, deploying your own still means learning Replicate's tooling and accepting its runtime conventions.

Who should pick this

Our pick when the workload is generative — images, video, audio, speech-to-text — and the traffic is intermittent or unpredictable. Product teams adding an AI feature to an existing app get the fastest time-to-working-API we know of: no containers to build for public models, no GPU quota wrangling, no idle bill. Researchers and hackers get access to hardware (H100, multi-GPU) that would be annoying to provision for a weekend.

We would pick something else in three situations. If you are serving sustained production traffic at high utilization, price out warm dedicated instances — /benchmarks-style arithmetic on your real duty cycle usually favors them. If you must process EU personal data or need SOC 2 evidence for a vendor review, Replicate is not there yet. And if you need guaranteed latency for user-facing requests, cold starts plus opaque routing are disqualifying without a warm-deployment strategy, which changes the cost picture.

One piece of standing advice: model the Cloudflare acquisition into your risk register. Acquired platforms get re-platformed, repriced, or merged, and we have watched that pattern often enough to say it out loud even when the current product is good. Today it is good — an unusually broad catalog, honest per-second billing, and a developer experience that delivers on its core promise.

Key takeaways

Hosted inference API with thousands of community models plus custom models packaged with Cog; pay per second of GPU time or per output unit depending on the model.
Hardware spans T4 ($0.81/hr) through H100 ($5.49/hr) and multi-GPU configurations up to 8x H100 ($43.92/hr), with scale-to-zero on public models and no idle charges.
New accounts since July 2025 are billed via prepaid credit: valid one year, non-refundable, runs halt at zero balance.
"Free" means a limited, undisclosed one-time allowance on select models — a trial, not a recurring free tier.
US-only infrastructure with no region pinning, no published uptime SLA, and no SOC 2/ISO certifications on record; weak fit for EU personal data.
Best for spiky generative workloads and prototypes; a poor fit for sustained high-utilization serving, where warm dedicated capacity is cheaper.

Sources

Pros

Thousands of callable models with one HTTP request; custom models via Cog containers
Per-second billing with scale-to-zero: no idle charges on public models, failed runs free
Hardware range from T4 to H100 and multi-GPU configurations up to 8x H100
Model version pinning keeps community updates from silently changing outputs
Strong SDKs, webhooks, and a web playground for evaluating models before coding

Cons

Per-second GPU pricing gets expensive at sustained high utilization
US-only infrastructure: no region selection, no EU data residency
No published uptime SLA and no SOC 2/ISO certifications on record
Cold starts on large models can run to minutes
Prepaid credit model: credits expire, are non-refundable, and runs halt at zero balance