HappyOyster 1.0

Not yet independently verified. The 17 September 2026 date is Model Studio's API listing; the HappyOyster product was first unveiled in April 2026 (Alibaba Cloud blog, 36Kr). No independent coverage of the September API release was found, and Acting is an invitational preview. We will update this when it can be confirmed, and remove this note.

HappyOyster 1.0 is Alibaba's real-time interactive world model, which generates a live, explorable video world from a text prompt and a first-frame image. Alibaba Cloud Model Studio lists three API models released on 17 September 2026: happyoyster-1.0-adventure, happyoyster-1.0-directing and happyoyster-1.0-acting. Only Adventure is listed for the China (Beijing) region, and the pricing page marks Acting as an invitational preview. HappyOyster itself is older: Alibaba's Alibaba Token Hub (ATH) business group unveiled it as a consumer product with early access in April 2026.

The three modes are deployed separately. Adventure lets a user explore a world with direction, camera and action commands. Directing drives a scene from text instructions or a structured script, with pause, rewind and resume, for interactive drama and film previsualisation. Acting generates a character in a chosen persona, in portrait 9:16 by default, for role-play and video-call-style companions. Integration uses server-side REST APIs to create worlds and issue short-lived credentials, and client SDKs for Android, iOS and the web that stream the result over a real-time audio and video channel. Each world is bound to the model that created it.

Singapore pricing is per call and per second of output video. Adventure costs $0.0077 per world created and $0.0308 per second at 480p. Directing costs $0.3113 per world created without a reference image or $0.0077 with one, $0.0154 per instruction, and $0.0539 per second at 480p or $0.0858 at 720p. Acting costs $0.0077 per world or instruction and $0.0308 per second at 480p or $0.0462 at 720p. Failed requests are not charged and there is no free quota. The models are proprietary and hosted only, and no parameter counts are published.