Access

AI video answer

How should a multi-model video API handle fallbacks?

Short answer

Fallbacks should run only for approved retryable failures, keep the same idempotent job, respect the cost ceiling, and record every attempted route. Invalid input and policy blocks should not retry another model.

01

Separate retryable from final errors

Provider unavailability, load shedding, or a temporary internal error may justify a delayed retry or approved fallback. Invalid media, unsupported settings, or a blocked prompt should return a clear final error.

Runway's task-failure guide makes the same distinction between temporary provider failures and input errors that should not be retried.

02

Protect cost and duplicate work

Use one idempotency key and one parent job. Before each fallback, estimate the next route and verify that the remaining cost allowance can cover it.

Do not launch several paid routes at once unless the caller explicitly approved parallel generation and its maximum total cost.

03

Show the full attempt trail

Store the route, upstream task ID, timestamps, failure class, charge, retry delay, and final result for every attempt.

The public result can stay simple while the attempt log gives support and billing teams the evidence they need.

Related guide

Get the full context.

Multi-model video API for routing, fallbacks, and cost control

Use one multi-model video API for capability routing, approved fallbacks, cost caps, normalized jobs, webhooks, and results across five AI video models.

Read the multi model video API guide

Current prices

Compare each video route.

Open full pricing →
ModelCompared rateCre8vid price5 sec10 secSavings
MiniMax H3MiniMax$0.065$0.029$0.145$0.2955% less
Wan 3.0Alibaba$0.050$0.025$0.125$0.2550% less
Kling 3.0Kuaishou$0.084$0.039$0.195$0.3954% less
Seedance 2.5ByteDance$0.144$0.069$0.35$0.6952% less
Veo 3.1 Fast + audioGoogle$0.100$0.049$0.24$0.4951% less