Chinese Labs Own the Leaderboard, Google Omni Leaks Before I/O, and the Post-Sora Production Stack

News

Gemini 'Omni' Video Model Surfaces in the Wild Ahead of Google I/O 2026

Real users spotted a live 'Create with Gemini Omni' prompt in the Gemini app this week — complete with early demo clips — just days before Google I/O kicks off on May 19. Early outputs show strong editing capabilities (object swaps, chat-based edits) but raw generation fidelity still trails Seedance 2.0, suggesting Google is betting on workflow unification over benchmark wins at launch.

Alibaba Reveals It's Behind HappyHorse, the Mystery Model That Topped Every Video Benchmark

HappyHorse-1.0 appeared anonymously on Artificial Analysis on April 7, hit #1 in both T2V and I2V within 48 hours, then Alibaba's ATH AI Innovation Unit claimed it — the same stealth-drop playbook DeepSeek used. The broader picture is stark: four of the top five video models on the current Elo leaderboard are Chinese-built (HappyHorse, Seedance 2.0, SkyReels V4, Kling 3.0), with Veo 3.1 the lone Western entry in the top five.

OpenAI Officially Confirms Sora Shutdown — App Dead April 26, API Gone September 24

OpenAI's official help doc confirms the two-stage Sora wind-down: app already gone, API expiring September 24, 2026. The shutdown was driven by $1M/day compute costs, user retention collapse (66% drop in 3 months), and a strategic pivot toward coding tools and enterprise — a cautionary tale about building a standalone video product without a clear monetization loop.

Analysis

Gemini Omni Deep Dive: What the Leaked UI Strings Actually Tell Us

TestingCatalog's detailed breakdown of the Gemini Omni leak reveals the model will ship in tiered Flash and Pro variants, treat video generation as an agentic task (like Deep Research), and support in-chat editing including watermark removal and object replacement. If this is a true unified text-image-video model, it would be the first of its kind from any top-tier lab — a structural shift, not just an incremental update.

Sora's Downfall Signals Broader Problems With AI's Creative Utility

This sharp post-mortem from The Conversation argues Sora's failure wasn't just about compute costs — it exposed a 'counter-creative bias' where AI tools optimize for familiar, aesthetically safe outputs rather than genuine novelty, leading to rapid curiosity-driven adoption followed by sustained disengagement. Worth reading for anyone building consumer-facing generative tools: the engagement cliff is real.

Seedance 2.0: The Full Copyright Crisis Timeline — MPA, Disney, SAG-AFTRA, and Senate Letters

The Wikipedia entry on Seedance 2.0 is now a comprehensive record of the most significant AI copyright confrontation of 2026: cease-and-desist letters from Disney, Paramount, Warner Bros., Netflix, and Sony, MPA calling infringement 'a feature not a bug,' and U.S. senators demanding a full shutdown. The legal status of the model's global CapCut rollout remains unresolved — a live risk for any production pipeline that touches it.

The 'Runway vs Pika' Frame Is Dead: Why the Real Story Is Chinese Labs Inside Western Platforms

This analysis makes a point that keeps getting overlooked: Runway already hosts Seedance 2.0 as a first-party model, meaning the US-vs-China framing of the video AI race is architecturally blurred. Five of the top ten Elo-ranked models are Chinese-built, and the 'multi-tool meta' — mixing Runway for coherence, Kling for photorealism, Pika for social effects — is now the actual production standard for serious video builders.

Tools

A Founder's Production Guide to Text-to-Video in 2026: Which Models Actually Ship

Fresh from Segmind (published in the last 24 hours), this is the most tactically useful thing in your feed today: real production data from millions of monthly video generations, with honest model routing — Seedance 2.0 Fast for high-volume ad variants, Veo 3/3.1 for cinematic work, Wan 2.2 for controlled image-to-video. The key insight is that text-to-video pricing has dropped ~5x in 12 months, and the winning workflow is almost always generate-a-still-first, then animate.