The model index
Text, audio, speech, avatars and video are billed five different ways. We convert all of them to one number — cost per minute and per hour of live conversation — then score each model twice, once for hiring and once for general work.
Every figure carries its source. Nothing appears as a bare claim.
OpenAI
The strongest general model for hiring work — but only worth its price on the judgement-heavy steps, not on parsing.
Long-horizon reasoning and the steadiest instruction-following in the catalogue.
Conversational video with perception — the avatar can see the person it is talking to.
The best Greek in our harness, and the only model here with speaker attribution.
$0.005 a minute, 300ms end to end. The price disruption is the product.
The only model in our harness that does not degrade under background noise.
The most expressive synthetic voice available, and the best voice cloning.
Steerable voice — you describe the delivery in a prompt rather than tuning sliders.
The most polished stock-avatar library, with real-time streaming on top.
Text-to-video with natively synchronised audio — one pass, not a dub.
The most physically coherent generated video — objects keep obeying physics across a shot.
Pre-rendered only — the enterprise standard for produced training video.
The model that started it all — and the cost floor everything is measured against.
Methodology
Cost per hour assumes a 150 wpm speaking pace and a 40% AI speaking share — the shape of a screening interview, where the model asks and the candidate answers. For token-billed models it also assumes three turns a minute against a cached prompt prefix. Hover any figure to see its full derivation.
One integration for text, voice, avatars and video — swap models without rewriting the pipeline.
