What VideoRouter is

VideoRouter is one API for all models and inference providers, with a focus on video and image generation models — OpenAI, Anthropic, Google, DeepSeek, xAI, and a long tail of open-weight and open video/image/audio models. Point your existing SDK at videorouter.sh/api, swap the key, and call any model by slug — for text/chat, let model:"auto" find the cheapest option for you, without sacrificing quality.

We built it as a genuinely cheaper, more transparent alternative to the aggregator model OpenRouter proved out — same one-API convenience, undercutting on the fee that actually matters.

The pricing thesis

Inference is billed at provider list price plus a flat 2% platform fee — across chat completions, embeddings, batch, audio (speech and transcription), and image and video generation — itemized as its own line — never blended into a bigger number. Depositing funds carries a separate, disclosed Stripe fee, the same honest split any aggregator's unit economics actually require. See Pricing & billing for the full breakdown.

Most models we serve direct from the provider's own API — no third-party markup baked into the price before our fee is even applied. A small long tail without a direct provider key is served through a backstop transport instead; see Models for how that's disclosed per model.

Availability, not just price

A cheap API that's down when you need it isn't cheap. Every request resolves to an ordered list of candidates — other providers of the same model, then other models entirely — so a single upstream outage rarely becomes your outage. See Model fallbacks and Provider selection.

Transparency by default

Every response carries headers explaining exactly what was routed, served, and billed — no opaque "trust us." Every request gets a per-trace drill-down at /logs, and account-wide spend/savings/latency at /usage. We publish what we log (and don't) at Data privacy & logging.

What's not built yet

In the interest of not overselling: there's no bring-your-own-key support, no contractual numeric uptime SLA (nobody in this market publishes one), and no reasoning-effort / extended-thinking token controls yet — those params are accepted but not specially billed or forwarded. Image input is supported; audio, PDF, and video input are not — see Vision.