Jev Daily

HeyGen's Decision 1 lands on OpenRouter at 93 ms median

Up to 400 calibrated probabilities per request across video, audio, images and text, with the latency and accuracy figures coming from HeyGen.

Decision 1 from HeyGen is live on OpenRouter. It takes questions about video, audio, images and text in a single request and returns a calibrated probability for each answer, up to 400 questions per call.

Models on OpenRouter
@ModelsOnOR
X
Ask questions about video, audio, images, and text in one request, and get a calibrated probability for each answer in ~100 ms.
Nov 20, 2026 · View on X

The account posting the listing puts the round trip at roughly 100 ms. OpenRouter's own post is more specific and attributes the figure to the vendor, quoting a median latency of about 93 ms on the Turbo variant.

The pricing

Launch pricing is 50% off through Dec 1. Turbo is $0.042 per million input tokens, the full Decision 1 model is $0.07 per million input tokens, and output is free. Output being free matters more than it sounds for this shape of model, because a 400-question batch returns 400 probabilities and you pay for none of them.

The accuracy claims

OpenRouter says the clearest leads are on audio tasks, citing speech emotion at 84.5 versus 32.4 and music genre at 88.1 versus 65.6. Those numbers come from HeyGen, and neither post says what the second number in each pair is. No baseline model is named, no dataset is named, and no eval harness is linked. Two score pairs without a named comparator are a marketing figure, not a result you can reproduce.

The same goes for the 93 ms. It is a median, the posts do not say over what payload, and video and audio inputs are not all the same size. A median on short clips tells you very little about what a selfie video costs you at the tail.

What it is being used for

OpenRouter lists moderation, identity checks, routing and QA as existing uses. The example it gives is an identity flow, checking a selfie video against an ID photo and asking "same person?" and "live capture?" in the same call, each returning its own probability. That is the real argument for the multi question format. Two independent judgements on one piece of media that would otherwise be two round trips.

For people building on it, treat the audio scores as unverified until someone outside HeyGen runs them. The thing worth testing first is the 400 question batch, because the per call economics only work if accuracy holds as you pack more questions into one request, and nothing in either post addresses that.

OpenRouter
@OpenRouter
X
The numbers behind it, from HeyGen: ~93 ms median latency on Turbo, and the clearest leads on audio tasks like speech emotion (84.5 vs 32.4) and music genre (88.1 vs 65.6).
Nov 20, 2026 · View on X

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.