Measured video performance

Video generation performance

Measured completion time, reliability, and cost for asynchronous video models served directly through the attested TrustedRouter gateway.

Generate video Text leaderboard

Last updated 2026-08-04T22:30:21Z
Video jobs are asynchronous, so token throughput and TTFT are the wrong measurements. This leaderboard uses end-to-end completion time, successful completion rate, exact customer cost, and cost per generated second from a rolling video benchmark set of up to 5,000 jobs. Rankings use successful completion first, then p50 completion time and cost. Prompt, reference media, generated bytes, download URLs, API keys, and workspace IDs are never stored in benchmark rows. Video traffic goes directly from the attested gateway to the named provider and never passes through OpenRouter.

Direct providers

Actual provider paths, not model publishers. A model offered through multiple direct providers receives one row per provider. Configured routes appear before their first completed measurement.

#ProviderModelsCompletedp50 totalp95 totalp50 cost / output secJobs
alibaba 1 Awaiting sample 0
atlas-cloud 1 Awaiting sample 0
google-ai-studio 2 Awaiting sample 0
grok 1 Awaiting sample 0
kling 2 Awaiting sample 0
ltx 2 Awaiting sample 0
minimax 1 Awaiting sample 0
openai 2 Awaiting sample 0
runway 1 Awaiting sample 0
venice 16 Awaiting sample 0

Video models

Operational measurements are kept separate from visual quality. A fixed public prompt suite and blind quality results will appear only after each model has enough comparable runs.

#ModelProviderInputsAudio Completedp50 totalp95 total p50 time / output secp50 job costp50 cost / output secJobs
alibaba/wan-2.7 venice Awaiting sample 0
alibaba/wan-2.7 alibaba Awaiting sample 0
bytedance/seedance-2.0 venice Awaiting sample 0
bytedance/seedance-2.0-fast venice Awaiting sample 0
google/gemini-omni-flash venice Awaiting sample 0
google/veo-3.1 google-ai-studio Awaiting sample 0
google/veo-3.1 venice Awaiting sample 0
google/veo-3.1-fast venice Awaiting sample 0
google/veo-3.1-fast google-ai-studio Awaiting sample 0
kling/o3-pro kling Awaiting sample 0
kling/o3-pro venice Awaiting sample 0
kling/v3-pro kling Awaiting sample 0
kling/v3-pro venice Awaiting sample 0
lightricks/ltx-2.3 ltx Awaiting sample 0
lightricks/ltx-2.3 venice Awaiting sample 0
lightricks/ltx-2.3-fast venice Awaiting sample 0
lightricks/ltx-2.3-fast ltx Awaiting sample 0
minimax/hailuo-3 minimax Awaiting sample 0
minimax/hailuo-3 venice Awaiting sample 0
minimax/hailuo-3 atlas-cloud Awaiting sample 0
openai/sora-2 venice Awaiting sample 0
openai/sora-2 openai Awaiting sample 0
openai/sora-2-pro openai Awaiting sample 0
openai/sora-2-pro venice Awaiting sample 0
pixverse/c1 venice Awaiting sample 0
runway/gen-4.5 runway Awaiting sample 0
runway/gen-4.5 venice Awaiting sample 0
shengshu/vidu-q3 venice Awaiting sample 0
x-ai/grok-imagine-video grok Awaiting sample 0
Next measurement layer

Quality without hand waving.

The quality suite will hold prompt, duration, aspect ratio, resolution, and audio settings constant. It will report prompt adherence, motion consistency, visual artifacts, identity consistency, text rendering, and audio synchronization separately.

Blind human preference and machine-judge scores will remain separate from reliability and cost. TrustedRouter will publish the prompts and replay metadata needed to reproduce every score.

Workspace access

Sign in

Choose a sign in method to access your TrustedRouter workspace.

By signing in you agree to the terms of service and privacy policy.