Running SDXL on RunPod Serverless — What It Actually Costs

The setup

RunPod serverless lets you deploy a Docker container to their GPU fleet, and you only pay for actual GPU seconds. No cold-boot penalty for warm workers. Ideal for spiky image workloads.

The math (mid-2026)

• RTX 4090 serverless: $0.00044 / second
• Typical SDXL 1024 generation: ~7 seconds wall clock
• Cost per image: $0.003

Add ~10 % for network egress and storage. Call it $0.0035 per image at cost.

Compare to hosted APIs

• Stability API (SD3.5): $0.065 / image
• Fal.ai SDXL: $0.02 / image
• Replicate SDXL: $0.007 / image
• RunPod your own worker: $0.003 / image

Self-hosting on RunPod is 2–20× cheaper than the hosted competition. That's why Imagoat runs its own worker instead of paying per-image at Replicate/Fal.

The catch

Cheaper only if you actually can operate it:

• You need a working Docker image (~1 day to get right)
• You need to handle safetensors download at boot (or bake it in — 6 GB per image)
• You need retry / timeout / error handling logic
• You need monitoring
• You need to keep the container updated as models evolve

For most people, paying 2× the wholesale cost via a hosted API is well worth avoiding the ops burden.

Where the break-even sits

Roughly 10,000 generations / month — below that, use a hosted API. Above that, self-hosting pays for the ops effort.

What Imagoat exposes

Imagoat handles the ops for you and passes RunPod pricing through in the credit system. Each SDXL Standard image is 1 credit — priced so heavy users get near-cost economics without operating anything.