Lumen opens Mirage 4 world model to developers

This is a work of fiction — an imagined dispatch from Vardis, a made-up world. None of the people, places or events in this story are real.
Lumen has put its Mirage 4 world model into the Halo API, letting developers build interactive, explorable worlds that respond to keyboard controls and typed instructions in real time. Mirage 4 became generally available in the Halo API on 19 November as models/mirage-4, through the Interactions API and Lumen Studio. It streams a continuous world in 720p resolution at 24 frames a second, with synchronised native audio, and developers can steer it with WASD and arrow keys or change the scene with a live text prompt.
Pricing is set at $1.50 per million input tokens, $9 per million text output tokens and $17.50 per million video output tokens. Lumen puts that at about $0.10 per second of streamed video, audio included, or roughly $6 a minute and $360 an hour.
From research preview to developer tool
Lumen Labs announced Mirage 4 on 27 October as a research preview for trusted testers, describing it as a real-time world model with native audio. The model generates music, ambient sound and dialogue together with the world, and its characters can talk and react as the scene unfolds. It is built on Halo Omni Flash, with Halo models supplying reasoning and steering in real time.
Consumers reached it first through Project Mirage, which began rolling out to eligible Lumen Prime subscribers on 18 November, with Lumen Pro members offered a place on a waitlist. The developer route through the Halo API followed a day later, on 19 November, and the pricing was published on Lumen's Halo API pricing page on 20 November.
How a session works
A developer starts a world from a text prompt, a single 16:9 reference image, or both. The model then streams the world continuously over WebRTC for browser apps or WebSocket for server code, using single-use ephemeral tokens for each session so that API keys are not exposed on the client. Audio is delivered as 48 kHz stereo, generated in the same pass as the picture rather than added afterwards.
Users can edit the world while it runs, changing the weather, adding a character or redirecting the action with a new prompt. Lumen's documentation warns that visual stability can drift after 60 to 90 seconds, and that looking away from an object for eight seconds or more, or turning a full circle, can move or reset landmarks, doors and buttons.
Short sessions, real costs
Each session is capped at 120 seconds and cannot be resumed, so a full session costs about $12.10 at the published rates. There is no pause mode, and the frame rate cannot be changed. Every output carries an imperceptible LumenMark watermark. Lumen Labs still describes Mirage 4 as an early, experimental model. Complex physical interactions, audio accuracy, legible text and long-term consistency remain open problems, and the model stays coherent over short sessions but can drift in longer runs or after several complex prompt edits.
Why it matters: world models are moving from video generation towards simulation environments for training robots and agents, and wiring Halo in as a planner makes that more concrete. At about $0.10 a second, Mirage 4 suits prototypes, design visualisation and short agent or robotics training sessions. A continuously running game server built on it would still be expensive: an hour of play costs about $363, spread over at least 30 separate worlds, none of which can be resumed once its session ends.