ByteDance's Seedance 2.5 opened its API on July 16 via BytePlus, and the headline feature is native 30-second generation without segment stitching — a genuine production leap. The model also accepts up to 50 reference inputs and bakes audio into the same latent space as video, meaning no post-sync workarounds. Still in global enterprise beta, and a copyright risk flag on training data is unresolved, so factor that in before going all-in commercially.
Google quietly launched Veo 3.1 Lite this week, priced at less than half of Veo 3.1 Fast while matching its generation speed. The play here is clearly high-volume production pipelines — creators running batch jobs will feel this immediately. Native portrait 9:16 and integrated audio remain available across the Veo 3.1 family.
TikTok has formalized a 4-tier AI disclosure system with C2PA auto-detection, and the enforcement teeth are real: unlabeled AI content depicting realistic people gets Creator Fund earnings zeroed for 14 days, even after retroactive labeling. AI-assisted text like scripts and hashtags is exempt. If you're posting faceless AI video on TikTok, label it on upload — there's no fixing it retroactively without a two-week income hit.
Kling AI raised $2 billion in July and is spinning out of Kuaishou as an independent entity, signaling serious long-term infrastructure investment. Kling 3.0 and 3.5 are already sitting at the top of the llm-stats video leaderboard with an arena score of 1934, and 3.5 adds prompt-less motion refinement from visual cues. At roughly $0.084 per second for standard tier, the pricing is competitive given the benchmark position.
Worth stating plainly for anyone still building around it: OpenAI shut down Sora on April 26, citing compute costs. The Sora 2 API goes dark completely on September 24. OpenAI has no announced video replacement in the near term.
MiniMax's Hailuo 2.3 earned the top Physics simulation ranking on WorldModelBench this month — best-in-class for realistic object dynamics and motion physics. This is a meaningful differentiator for product visualization, food and liquid shots, and anything where Newtonian accuracy matters. Pricing remains accessible compared to Veo and Runway.
Alibaba has confirmed Wan 2.7 is coming, promising better visual quality, improved audio, enhanced motion dynamics, and 9-grid image-to-video editing. However, there's no release date, and Wan 2.5 is API-only — meaning Wan 2.2 remains your ceiling for local ComfyUI inference. The Wan2.2-T2V-A14B-GGUF quantized build is still trending on Hugging Face this week.
As of July 10, Runway's API now supports negative prompts up to 1,000 characters across Veo 3, Veo 3.1, and Veo 3.1 Fast. This is a workflow unlock for anyone who has been fighting prompt drift or unwanted artifacts in Veo outputs — particularly useful for character consistency across longer pipelines.
Lightricks pushed a major update to the LTX Trainer this month, adding support for video LoRA, audio conditioning, and cross-modal IC-LoRA fine-tuning all in one framework. LTX-2.3 already has 18 million-plus Hugging Face downloads since March. The active "AI on the Lot 2026" ComfyUI workflow series is building around it, and the community QwenVL Prompt Enhancer workflow pairs well with it for structured prompt generation.
Comfy Desktop finished its rollout on June 8 and now functions as a multi-install manager similar to Stability Matrix — you can run and switch between multiple ComfyUI environments from one interface. On the performance side, the NVFP4 quantization plus async offload plus pinned memory trio is delivering meaningful VRAM relief on NVIDIA GPUs. New partner nodes this month include Nano Banana 2 Lite and GPT-5.5 Pro.
