// tools/api-latency-comparison

Inference API Latency Comparison

Warm generation times across 6 providers — this is the latency your users see once the model is loaded, not the cold start. For cold start data, see the Cold Start Benchmarks.

Single image · 1024×1024 · warm requests only · 2026-06-01

4-step distilled Flux · 4 steps

Fastest for Flux Schnell: Runware800ms median · 1.8s P95
Fastest raw inference. Purpose-built high-throughput stack.
fastest
RunwareInference API
pricing →
P50 — median request
800ms
P95 — worst 5% of requests
1.8s
Fastest raw inference. Purpose-built high-throughput stack.
Optimized CUDA kernels, lowest latency on market for Flux Schnell.
fal.aiInference API
pricing →
P50 — median request
1.8s
P95 — worst 5% of requests
3.5s
Widest model coverage, consistent warm fleet.
Novita AIInference API
pricing →
P50 — median request
2.0s
P95 — worst 5% of requests
4.0s
Competitive on Flux and SDXL pricing.
Together AIInference API
pricing →
P50 — median request
2.5s
P95 — worst 5% of requests
5.0s
Competitive pricing, free tier for new devs.
Shared fleet — usually warm for popular models.
ReplicateInference API
pricing →
P50 — median request
2.5s
P95 — worst 5% of requests
5.0s
Largest model catalog including community models.
RunflowComfyUI ManagedComfyUI overhead included
pricing →
P50 — median request
3.5s
P95 — worst 5% of requests
7.0s
Enterprise SLA, custom ComfyUI workflows, LoRA hot-swapping, multi-step pipelines.
Single-model ComfyUI workflow. Complex pipelines add ~0s marginal overhead.
Runflow runs ComfyUI workflows — latency includes workflow orchestration overhead. For complex multi-step pipelines (LoRA + upscale + post-process), this overhead becomes proportionally negligible.

Flux Schnell — side-by-side comparison

ProviderP50P95Type
Runware800ms1.8sInference API
fal.ai1.8s3.5sInference API
Novita AI2.0s4.0sInference API
Together AI2.5s5.0sInference API
Replicate2.5s5.0sInference API
Runflow3.5s7.0sComfyUI Managed

Reading these numbers

  • P50 is the median — half of requests are faster, half are slower. This is your "typical" user experience.
  • P95 is what 95% of requests are under. Your slowest 5% users see this. Design SLAs around P95.
  • Runflow is a ComfyUI platform — latency includes workflow scheduling overhead. It excels for complex multi-step pipelines (LoRA + upscale + post-processing) where that overhead is proportionally negligible.
  • • For latency-sensitive UX (under 3s): Runware or fal.ai for Flux Schnell. Both <2s median warm.
  • • For background/batch jobs: latency is irrelevant — optimize for cost per image instead.
Estimates from provider documentation and community benchmarks — not independent lab measurements · Cold start benchmarks →