xAI Launches Imagine Multi-Camera Video Agent
xAI announced the launch of Grok Imagine Video 1.5 agent, integrating programmable agent workflows with image and video models to plan narratives and stitch together multi-camera shots. The driving layer has been upgraded to the latest Image 2.0, with official claims of higher image quality, more complete narratives, and stronger continuity between characters and scenes across multiple shots.
Video 1.5 is not an entirely new foundation. The model was fully opened in June via the Imagine API, with an accelerated version called Fast: generating 6 seconds of 720p video in about 25 seconds, compared to over 40 seconds for similar models. By the end of July, it added pure text-to-video generation, image and sound references, and native 1080p. Up to seven elements can be locked in reference images, and sound references are used to fix the same face and voice. Sound effects, ambient sounds, and dialogues can be generated in a single inference with the visuals.
The agent layer aims to address the issue of insufficient single 15-second segments: first breaking down story beats, then generating by shots, and finally stitching multiple segments into one. The official demonstration emphasized that "the continuity is better when connecting multiple shots." Product manager Andrew Milich described it as the planning capability of coding agents combined with the generative capabilities of image and video models, aimed at creating narrative with judgment rather than producing a random short clip all at once.
Image 2.0 handles 1K or 2K still frames, emphasizing layout, composition, and instruction following, with partition editing and segmentation tools. The video API is priced based on duration and resolution, with public rates approximately $0.08 per second for 480p, $0.14 per second for 720p, and $0.25 per second for 1080p, plus fees for reference image inputs. Watermarks on the consumer side cannot be removed. Access is available at grok.com/imagine as well as on iOS and Android.
In comparison with competitors, Video 1.5 had previously ranked high in the image-to-video category during its preview period, and subsequent public evaluations showed narrowing gaps with models like Gemini Omni Flash. OpenAI's Sora had once contracted due to resource and business model issues, while Google’s Veo, Seedance, and Tongyi Wanxiang continue to compete in the same short video space for creator workflows. xAI emphasizes the difference in being an "agent director" rather than just a single model scoring.
In terms of market mechanisms, buyers are creators with high demands for short dramas, advertisements, and character consistency, while sellers are generating APIs billed by the second. Funding has shifted from one-time prompt trials to reusable character references, voice locks, and project-level multi-agent parallelism. Beneficiaries are platforms that can sell 15-second slices as continuous narratives; those under pressure are old tools that only sell single shots without face and voice locking. The event-driven clarity is that the model is already available, but what is missing is the scheduling layer that can connect shots into stories.
On a supplementary note, the official has not disclosed the maximum number of shots an agent can plan at once, the longest duration of a completed film, or the interface boundaries between Image 2.0 and Video 1.5 in reference video editing; third-party tools have already begun to integrate Video 1.5 and Image 2.0.
Source: Public Information
ABAB AI Insight
xAI's approach to video is similar to that of large models: first embedding the generator into Grok, then refining the workflow into a product. Imagine has evolved from still frame Aurora to Video 1.0, 1.5, and now Image 2.0, amidst safety controversies and deepfake criticisms. Andrew Milich refers to this iteration as an "agent," acknowledging that single model generation does not solve the real payment issues for creators—not just animating a single image, but ensuring the same character remains consistent in the next shot. Historical comparisons show that Runway first sold Gen-2 before offering multi-camera tools, while Pika initially sold gameplay before addressing consistency, ultimately getting stuck on "continuing a segment leads to face distortion."
Capital mobilization points towards API billing and subscription stratification. At $0.25 per second for 1080p, long videos become predictable cloud costs, with SuperGrok at a higher level first obtaining reference images and voice locks before scaling to all users. The motivation is to transform Grok from a chat window into a generative studio: users no longer just consume text tokens but rather consume image production capacity measured in seconds. This aligns with CoreWeave's competition for Rubin cabinets—upstream competing for computing density, downstream competing for pricing power per second of video.
Analogous examples include Adobe embedding Firefly into Premiere, OpenAI embedding Sora into ChatGPT but retracting on commercialization, and Google integrating Veo into the Gemini workflow. xAI is transitioning from "model release" to "controlling the creative entry point." Whoever masters shot planning will not have to fight for every shot; the agent upgrades prompts into a production process.
Structural judgment belongs to technological substitution. The mechanism is: as the quality gap of single shots narrows, competition shifts from pixels to continuity. A scheduling layer that can lock faces, voices, and scenes will replace a significant amount of manual alignment on editing tables. Generated videos will no longer be sold based on "a single prompt" but rather on "whether a scene can hold together." Consequently, pricing power shifts from model scoring rankings to platforms that can manage multiple shots as a single project.
ABAB News · Cognitive Laws
Single shots focus on image quality, while multi-shots focus on identity.
Model generates visuals, while agents generate structure.
Locking characters first ensures locking duration.