OpenAI pulls the Videos API and all Sora 2 model aliases on September 24, 2026, completing the shutdown that started when the Sora app went dark on April 26. The deprecation table leaves the "recommended replacement" column blank — there is no successor endpoint inside OpenAI. If you have unexported generations, sora.chatgpt.com/sunset is the only door, and OpenAI permanently deletes the data after the final export window.
Video Rebirth dropped a distillation LoRA on September 19 that takes H3 sampling from 49 steps down to 8, with ComfyUI ports already circulating. That's roughly a 6x throughput win on the most-downloaded open video model on Hugging Face right now. Combined with the vLLM-Omni and FastVideo FastH3 path — 10.1 seconds of video plus audio rendered in about 8.7 seconds — local H3 is now genuinely faster than playback.
The September 15 release adds native FastVideo FastH3 support for text-to-video and image-to-video with native audio, plus Marigold V2 for depth, normals and albedo, and YuE2 for music generation. On the graph side, Video Concatenate finally makes multishot stitching a first-class node and Generic Loops lets you iterate over lists without custom nodes. H3 video VAE now runs Comfy-Kitchen kernels, and YuE2 max song length jumped to 900 seconds.
UntMods released an H3 FaceSwap LoRA today, September 20, that swaps identity into a reference video from a single facial image, shipping with a ready-to-run ComfyUI integration. The Apache-2.0 license is the headline — that's permissive in a space where the H3 base weights themselves sit under a restrictive community license excluding the US, EU, UK and South Korea. Obvious consent and likeness landmines here, so gate it accordingly.
XGEN Labs published weights and code for an egocentric interactive model built on MiniMax H3 — you type a WASD camera path, feed reference images, and it generates first-person video plus synced audio with object interaction and character dialogue. Caveat: this release is a four-step bidirectional model, meaning the whole action sequence has to be known up front. The causal version that reacts key-by-key isn't out yet.
The 2026 sweep has removed 16 major AI channels holding roughly 35 million subscribers and 4.7 billion cumulative views. Three demonetizing categories are in force: generic template-based output, off-putting or distressing content, and AI personas discussing sensitive topics like health and finance. Creators with visible human direction and proper disclosure report no adverse action — the trigger is mass-production patterns, not the use of AI itself.
Lightricks' 22B open-weights audio-video model is at 1.61 million downloads and 4,428 likes on Hugging Face, trailing only MiniMax H3's 4.30 million. The distilled variant renders a 10-second clip from an image in about 6.8 seconds on NVIDIA superchips, with native multishot scenes, auto duration and 4K HDR. Free for orgs under 10 million dollars ARR; above that you negotiate.
Kuaishou's Kling v3 sits at 1,934, ahead of Alibaba's Happy Horse 1.0 at 1,816 and ByteDance's Seedance 2.0 Fast at 1,747. Veo 3.1 still owns synchronized dialogue with 48kHz speech, and Runway Gen-4.5 remains the pick when camera control matters more than leaderboard position. Seedance 2.5 is the model creators are calling the overall king — native 30-second 4K, up to 50 reference images — but access is still throttled.
Released September 17, RefMods packs reference photos into compact files tuned directly against MiniMax H3, with preview before you commit to a ComfyUI run. That's the missing middle between one-shot reference conditioning and a multi-hour LoRA train — the exact bottleneck for anyone running episodic or recurring-character content. Pair it with the HyperFlow 8-step LoRA and your iteration loop shrinks dramatically.
Wan 3.0 has been in paid public beta since August 6, generating 30-second clips in a single run and accepting documents as input, priced at 0.3 to 1.2 yuan per second by resolution. No weights exist. The last open-weights flagship in the family is Wan 2.2 from July 2025 under Apache-2.0 — 2.5, 2.6, 2.7 and 3.0 all shipped closed. If your workflow still depends on Wan locally, you're building on a frozen base.
