Replicate

“(model gallery hero — browsing interface over infrastructure value proposition)”

Leading with a model grid makes Replicate look like a directory. The developer who lands here from a "run Stable Diffusion API" search needs to see "zero infra, pay per second" before they see a browse experience they can get on Hugging Face for free.

“Run any open-source AI model with one API call. No GPUs to manage, no Docker to configure, no idle costs. Pay only for the seconds you use.”

“(per-second pricing shown without cost examples or monthly-spend estimator)”

Per-second billing sounds cheap until the developer realizes they have no mental model for what 10,000 image generations cost. Without a calculator or example invoices, the pricing feels unpredictable — which is the exact objection Replicate should be eliminating.

“Add worked examples: "1,000 SDXL images: ~$1.30. 10,000 Whisper transcriptions (avg 5 min): ~$4.50. See the full price list with cost-per-unit for every model."”

“(cold-start latency not addressed on the model pages or pricing)”

Open-source model inference on shared infra means cold starts. Developers who hit a 15-second first-request delay without warning will distrust latency claims and evaluate self-hosting instead.

“Be transparent: "First request: 5-15s cold start. Subsequent requests: <1s. Need always-on? Use dedicated endpoints — predictable latency, fixed hourly rate."”