Here's the briefing:
---
🎬 AI Video Intel — 2026-04-26 06:45
Model & Tool Releases
• Sora dies today. The Sora app and web experience shut down April 26, 2026. API stays live until Sep 24. OpenAI cited compute costs, declining MAU (1M → <500K), and strategic pivot to enterprise/coding. Disney killed its $1B stake after learning of the shutdown <1 hour before announcement. (OpenAI Help Center, Variety, TechCrunch)
• HappyHorse-1.0 tops all leaderboards. Alibaba's stealth entry (via Taotian Future Life Lab, led by ex-Kuaishou/Kling head Zhang Di) hit #1 on Artificial Analysis Video Arena at 1,389 Elo — 115 points above Seedance 2.0. Supports T2V, I2V, with and without native audio in a single forward pass. Phoneme-level lip-sync in 7 languages. API launches April 30. Open-source status conflicting — fal.ai says closed, some sources say open. Multiple fake fan sites already flagged by official X account. (CNBC, fal.ai, GIGAZINE)
• Wan 2.7 dropped late March. 27B params (14B active via MoE), Diffusion Transformer + Flow Matching. New: first/last frame control, 9-grid multi-scene generation (3×3 image grid → continuous video with transitions), subject + voice cloning from reference clips. 1080p up to 15s. Cloud-only for now — open weights expected 4–8 weeks per Alibaba's established pattern. (ComputerTech, WaveSpeedAI)
• Hailuo 2.3 (MiniMax) rolled out globally. Better physical action rendering, anime/illustration style support (ink wash, game CG), improved micro-expressions. Video Agent rebranded to "Media Agent" for full multimodal creation. 2.3 Fast variant cuts batch costs ~50%. Available via API on fal.ai. (MiniMax, fal.ai)
• Runway added Seedance 2.0 to its platform (April 7) for Unlimited/Enterprise plans outside the US. Gen-4.5 continues as their flagship with native audio and multi-shot one-minute generation. (Runway Changelog)
Platform Signals
• TikTok: C2PA Content Credentials now auto-detect synthetic media even without creator disclosure. AI content isn't suppressed if labeled — but users can now filter how much AIGC they see. Script writing and hashtag generation remain exempt from labeling. AI influencers are permitted with clear labeling. (Storrito)
• Instagram Reels: The Originality Score is now a real distribution lever. It detects recycled clips, duplicates, low-effort reposts — and AI-generated content with unique visuals + original voiceover + custom captions scores HIGH because each video is genuinely novel. This is the key insight: AI-gen isn't penalized, it's rewarded for being original-per-piece. (FlowShorts)
• Meta (IG/FB): Automatically displays "AI Info" / "Made with AI" when C2PA metadata is detected in uploads. (Influencer Marketing Hub)
• X/Twitter: Effective March 3, 2026 — posting AI-generated armed conflict videos without disclosure triggers 90-day Creator Revenue Sharing suspension. (X Help Center)
What's Working
• $61K/month AI influencer case study: Single repeatable UGC format — same AI face, same bedroom setup, same lighting. Swaps product in hand and updates script. Posts daily across dozens of offers. No filming, no brand negotiation. (@georgesttock on X)
• Rizz app: $15M+ revenue from creator distribution network generating 5B views across social. Paid $2M to faceless creators to build that engine. 7.5x return. (@fuckgrowth on X)
• Virtual influencer sponsorship rates: $2K–$15K per sponsored post at 100K+ followers in tech/fashion/fitness. IP licensing deals averaging $45K–$120K for large brand campaigns. (CommuniPass)
• AI-driven revenue optimization (pricing, churn prediction, subscriber retention) showing 35–60% earnings lift without new content. Reducing monthly churn by 5% compounds to 60–80% higher annual revenue. (FinancialContent, April 20)
Technique Watch
• Z-Image two-sampler pass gaining traction in ComfyUI video workflows: render structure tiny (288×288) in first sampler, 6× latent upscale, then second sampler for detail. Much better quality, stays fast. Best for T2V. (markdkberry.com)
• Practical LoRA stack for video: BSF Head Swap (character consistency), Blending Fusion (object integration + relighting), Kijai Adapted Distill at -0.3 to 0.6 (fixes frozen-frame artifacts). Running on RTX 3060 12GB. (markdkberry.com)
• LTX 2.3 TextGenerateLTX2Prompt node auto-rewrites plain language into model-friendly prompts covering actions, visuals, and audio cues. Useful for batch workflows where manual prompt engineering doesn't scale. (RunComfy)
• Production pipeline pattern: QWEN Image Edit 2511 → Z-Image (low denoise for realism) → SeedVR2 4K upscale. Working res 480×201 (2.39:1 cinematic), upscaled 2× to 1920×1080. (markdkberry.com)
Worth Watching
• HappyHorse API (April 30) — if the Arena rankings hold and pricing is competitive, this immediately becomes the model to test for production pipelines. The Zhang Di/Kling pedigree is real. Watch for whether "open-source" actually materializes or if that was fake-site misinformation.
• Wan 2.7 open weights — expected within weeks. If the 9-grid multi-scene and subject cloning capabilities survive the open release, this is a game-changer for local ComfyUI pipelines. Keep an eye on the Wan-Video GitHub.
• Sora vacuum: With Sora dead and ~500K users displaced, expect a land-grab from Kling, Runway, and Hailuo over the next 2 weeks. Watch for aggressive free-tier promotions.
• DeepSeek V4 launched this week (Flash + Pro variants) — not video-specific, but their compute efficiency approach could show up in video models next. (Bloomberg)
• Instagram Originality Score is the sleeper policy change. If you're producing unique AI-gen content per piece (not repurposing the same clip), you're structurally advantaged over human creators who recycle formats. This is worth building around.