“OpenRouter, but for media” — and where the router stops and the proof layer starts

This is the comparison we take most seriously, and the one where we concede the most. OpenRouter is the product this entire category studies — our own strategy docs study it by name. It is a neutralrouting layer over hundreds of third-party models, with no first-party model of its own to protect, which is exactly what makes it more of a real competitor than any single model host. And in 2026 it moved into media: a unified image API across dozens of image models, and video generation across the top text-to-video and image-to-video models. So "OpenRouter for media" is no longer a hypothetical. It ships.

What OpenRouter does — precisely, and well

It's worth stating the pricing model exactly, because it's clean and we're not going to pretend otherwise. OpenRouter takes no markup on inference— you pay the underlying provider's token or per-unit price. It makes its money on a fee when you buy credits (5.5% with a $0.80 minimum on card payments, less on crypto), and its bring-your-own-key option is priced as a small percentage (5%) of what the request would have cost, with the first million requests a month waived. One balance, hundreds of models, honest pass-through economics. For a lot of teams that is the right answer, and now that it routes images and video, it's a right answer for media too.

Routing is where OpenRouter is strong and where we'd rather concede than exaggerate. The media difference starts after the bytes come back.

Where a router stops and a proof layer starts

A router's job ends when the response returns. That's the right design for text, where the output is ephemeral and the interesting thing is the tokens. Generative media is different: the output is an asset with a lifecycle — it has to be captured before a provider URL expires, stored somewhere you control, hashed so you can prove it later, and traced when it becomes the input to the next job. None of that is routing, and a layer built LLM-first doesn't reflexively do it:

Side by side, honestly

Routing across providers, made visible with three real renders from three different ones (fal, Runware, PiAPI). But the part that matters against a router is what happens to each output after the response returns — streamed into byte-exact custody with its own job id and sha256. The full one-prompt-through-five-providers proof is on the comparison overview.

Real pipeline outputs — model + provider named per tile; job ids and sha256 live in the showcase manifest, not in a caption.