A unified multimodal AI API gives applications one identity, policy, model catalog, observability layer, and bill across text, image, video, and audio. It should unify the control plane while preserving the correct execution model for each modality.
The INFRO view: INFRO is designed as one infrastructure and billing layer for text, image, video, and audio models while documenting their different request lifecycles honestly.
What should be unified
Authentication, projects, model identifiers, access policy, budgets, request IDs, error envelopes, usage records, and billing benefit from one vocabulary. This is the layer product and platform teams should not have to rebuild for every vendor.
A shared catalog also makes capabilities explicit: modality, inputs, limits, lifecycle, and current availability.
What should remain modality-specific
Text can stream tokens over seconds. Image jobs may return one or more assets. Video can run for minutes and needs durable status. Audio may be synchronous for speech or asynchronous for longer processing. Forcing all four into a chat-completions shape produces leaky abstractions.
Use synchronous responses where the work fits and asynchronous jobs with signed webhooks where it does not. The API can still share auth, status, errors, and usage accounting.
Design the job lifecycle
For asynchronous media, create a job, return a stable identifier, expose status, sign webhook deliveries, make handlers idempotent, and support safe polling. Record model, parameters, resolution or duration, cost, and asset expiry.
Plan for cancellation and duplicate events. A webhook can be delivered more than once, and a client timeout does not mean the media generation stopped.
Govern cost and safety across modalities
Token budgets are not enough. Image count, resolution, video seconds, frames, and audio duration need their own units and ceilings. Access policy should prevent development keys from accidentally invoking expensive media models.
Keep modality-aware safety checks and provenance requirements near the job layer so every application does not invent its own inconsistent controls.
Where INFRO fits
Text uses an OpenAI-compatible endpoint. Media work uses asynchronous jobs and signed webhooks where appropriate. The shared control plane supplies model access, routing, request identity, usage, and spend visibility.
Review the modality documentation, current model catalog, and pricing. Because INFRO is in private early access, confirm the exact media models and job features enabled for your account.
Frequently asked questions
Can every AI modality use the chat completions format?
Not cleanly. Text streaming, image assets, long-running video jobs, and audio workflows have different lifecycles. A good platform unifies control without hiding those differences.
What does INFRO unify across modalities?
INFRO unifies supported model access, identity, policy, usage, observability, and billing across text, image, video, and audio.
How are long-running INFRO media jobs delivered?
INFRO documentation describes asynchronous media jobs with signed webhooks; verify the current endpoint and model details before implementation.
