## Sources 1. [Cerebras Grew 92% - Investors Saw the Margins and Sold](https://awesomeagents.ai/news/cerebras-q1-2026-earnings-margin-crash/) 2. [Kling 3.0](https://awesomeagents.ai/models/kling-3-0/) 3. [Grok Imagine Video 1.5](https://awesomeagents.ai/models/grok-imagine-video/) 4. [Dreamina Seedance 2.0](https://awesomeagents.ai/models/seedance-2-0/) 5. [Wan 2.7](https://awesomeagents.ai/models/wan-2-7/) 6. [HappyHorse-1.0](https://awesomeagents.ai/models/happyhorse-1-0/) 7. [SkyReels V4](https://awesomeagents.ai/models/skyreels-v4/) 8. [Sora 2](https://awesomeagents.ai/models/sora-2/) 9. [Runway Gen-4.5](https://awesomeagents.ai/models/runway-gen-4-5/) 10. [Best AI Models for Video Generation - June 2026](https://awesomeagents.ai/capabilities/video-generation/) --- The following summary details the key concepts and developments from the provided sources as of June 2026, focusing on the competitive landscape of AI video generation and recent semiconductor industry news. ### **Best AI Models for Video Generation - June 2026** **Author: James Kowalski** [1] * **Market Leadership Shifting:** The AI video generation market is in a state of rapid flux, with top positions on leaderboards typically changing every 60–90 days [2, 3]. * **Top Visual Performance:** Alibaba-ATH’s **HappyHorse-1.0** currently holds the #1 spot for visual quality (without audio) on the Artificial Analysis Video Arena with an Elo of 1,290 [1, 4]. * **Top Audio Performance:** ByteDance’s **Dreamina Seedance 2.0** remains the leader for video generated with native audio, holding an Elo of 1,218 [1, 5]. * **OpenAI's Exit:** OpenAI is effectively exiting the consumer video space; the Sora 2 consumer app closed in April 2026, and the API is scheduled for sunset on September 24, 2026 [3, 5]. * **Emergence of xAI:** xAI entered the competitive tier in June 2026 with **Grok Imagine Video 1.5**, which is already showing strong benchmark potential [5, 6]. ### **Cerebras Grew 92% - Investors Saw the Margins and Sold** **Author: Daniel Okafor** [7] * **Revenue Growth vs. Margin Compression:** Cerebras reported a massive 92% year-over-year revenue growth for Q1 2026 ($193.4M), yet its stock fell approximately 20% due to a dramatic cut in gross margin guidance [7, 8]. * **Guidance Downturn:** While Q1 margins were 47%, guidance for Q2 dropped to 36–38%, far below industry leaders like Nvidia (70%+) [7]. * **Data Center Strategy:** Management attributes falling margins to the cost of simultaneously building their own data centers while renting capacity from customers to meet demand [9, 10]. * **OpenAI Relationship:** Cerebras has a $20 billion compute supply agreement with OpenAI, which currently runs GPT-5.4 on Cerebras infrastructure [11, 12]. * **Valuation Reset:** The stock’s collapse from its $350 IPO-day opening to around $182 represents a significant valuation reset as investors move away from expecting Nvidia-level margins [8, 10]. ### **Dreamina Seedance 2.0** **Author: James Kowalski** [13] * **Integrated Synthesis:** This model is notable for creating video and audio in a **single forward pass** using a dual-branch Diffusion Transformer architecture [13, 14]. * **Leader in Audio-Visual Quality:** It leads the "with audio" category in both image-to-video (Elo 1,194) and text-to-video (Elo 1,219) benchmarks [15, 16]. * **Multi-Shot Narrative Control:** Unlike many competitors, Seedance 2.0 can produce coherent scene cuts and transitions within a single 15-second generation [17, 18]. * **Global Availability and Risk:** Launched globally via fal.ai in April 2026, the model faces legal pressure from Hollywood studios and regulatory scrutiny from the US Senate regarding the generation of likenesses [15, 19]. * **Multimodal Flexibility:** The "reference-to-video" workflow is highly flexible, accepting up to 9 images, 3 video clips, and 3 audio files as inputs [20, 21]. ### **Grok Imagine Video 1.5** **Author: James Kowalski** [22] * **Top-Ranked Image-to-Video:** xAI's model holds the #1 spot on the Artificial Analysis Image-to-Video Arena (Elo 1,473 at launch) [22, 23]. * **Autoregressive Architecture:** Built on the "Aurora" engine, it uses an **autoregressive** approach (sequential frame generation) rather than diffusion, leading to more stable motion over its 15-second clips [22, 24]. * **Cost Efficiency:** At $0.14/s for 720p, it is significantly cheaper than competitors like Sora 2 Pro and Google's Veo 3.1, with audio included at no extra charge [25, 26]. * **Rapid Iteration:** The model moved from its initial August 2025 launch to its current top-tier status in roughly ten months, aided by a large feedback pool from over a billion user-generated videos [23, 27]. * **API Constraints:** While powerful, the API currently does not support text-to-video (only available via the consumer interface) and is capped at a 720p resolution for standard use [28, 29]. ### **HappyHorse-1.0** **Author: James Kowalski** [30] * **Surprise Leader:** This model debuted anonymously on leaderboards before being revealed as a product of Alibaba’s ATH Innovation Unit, led by former Kuaishou VP Zhang Di [30, 31]. * **Visual Quality King:** It ranks #1 in text-to-video (Elo 1,357) and image-to-video (Elo 1,402) for silent video [32, 33]. * **Unified Transformer:** It uses a 15-billion-parameter unified transformer architecture that processes text, image, video, and audio tokens in a single attention sequence [32]. * **Version 1.1 Improvements:** The 1.1 update added native multilingual lip-sync across seven languages and established 1080p as the standard output [34, 35]. * **Pricing:** Available via fal.ai at $0.14/s for 720p and $0.28/s for 1080p [34, 36]. ### **Kling 3.0** **Author: James Kowalski** [37] * **Native 4K Pioneer:** Kling 3.0 is the first commercial model to ship **native 4K at 60fps** without upscaling [37, 38]. * **Tiered Offerings:** The "Omni" (O3) variant adds advanced features like multi-shot storyboarding, vector-based camera path control, and "advanced reasoning" to plan scenes before rendering [39, 40]. * **Competitive Pricing:** The API is priced at $0.075/s for standard generations, making it one of the most affordable high-quality production options [41, 42]. * **Granular Controls:** Features like "Motion Brush" allow users to paint specific motion vectors onto up to six different elements in a static image [43, 44]. * **Global Standing:** It holds a top-three global position with an Elo of 1,251 for non-audio video [37, 45]. ### **Runway Gen-4.5** **Author: James Kowalski** [46] * **A2D Architecture:** Built on an "Autoregressive-to-Diffusion" architecture co-developed with NVIDIA, combining language-style scene understanding with diffusion visual quality [47, 48]. * **Production Focus:** It is highly regarded for its **Motion Brush 3.0** and "Camera Director" tools, which offer precise region-level animation and camera choreography [49, 50]. * **Strategic Partnerships:** Runway is the preferred API partner for **Adobe Firefly**, integrating Gen-4.5 directly into Adobe's creative ecosystem [48, 51]. * **Competitive Shift:** Though it held the #1 Elo spot at its December 2025 launch, it has since been surpassed by models like Seedance 2.0 and Kling 3.0 in blind preference tests [46, 52]. * **No Native Audio:** A major weakness is its lack of native audio generation, requiring sound to be added in post-production [51, 53]. ### **SkyReels V4** **Author: James Kowalski** [54] * **Joint Generation Differentiator:** Skywork AI’s V4 model uses a dual-stream system to generate 1080p video and synchronized audio simultaneously rather than sequentially [54, 55]. * **Unified Pipeline:** It is the first model to combine generation, inpainting, and editing into a single unified pipeline [54, 56]. * **Character Consistency:** It features "Grid Reference," allowing users to upload up to nine images to maintain consistent character features and style across a sequence [57, 58]. * **Benchmark Performance:** It briefly held the #1 spot in the "With Audio" category in March 2026 and currently sits at 4th place (Elo 1,104) [55, 59]. * **Affordability:** Priced at approximately $0.12/s with audio, it is significantly cheaper than Google's Veo 3.1 [59, 60]. ### **Sora 2** **Author: James Kowalski** [61] * **Physics Excellence:** Launched in late 2025, Sora 2 was noted for having the best physics simulation in the industry, correctly modeling complex cause-and-effect scenarios like fluid dynamics [61-63]. * **Tiered Access:** The model offered Sora 2 (720p) and Sora 2 Pro (up to 25s at 1080p) [64, 65]. * **Likeness Feature:** The consumer app included "Cameos," allowing users to insert their verified likeness into generated clips with granular consent controls [66, 67]. * **Strategic Discontinuation:** OpenAI is shutting down the product to reallocate compute resources toward enterprise and coding tools; the API sunset is September 2026 [68, 69]. * **Falling Behind:** By mid-2026, it had been surpassed by competitors in resolution (4K) and clip duration [65, 70]. ### **Wan 2.7** **Author: James Kowalski** [71] * **Open-Source Standard:** Released under **Apache 2.0**, Wan 2.7 is the leading open-weight video model, allowing for self-hosting and fine-tuning [71-73]. * **MoE Architecture:** It utilizes a 27-billion-parameter Mixture-of-Experts (MoE) architecture that activates only 14 billion parameters per pass for efficiency [74, 75]. * **Deterministic Control:** It features a "first-and-last-frame" (FLF2V) mode that allows users to provide starting and ending images to generate the transition between them [76, 77]. * **Thinking Mode:** This feature adds a pre-rendering planning pass to interpret complex prompts and reduce subject drift [78, 79]. * **Benchmark Strength:** It ranks 4th on the image-to-video (with audio) leaderboard (Elo 1,090), performing competitively with closed models at a fraction of the cost ($0.10/s via API) [71, 80].