Summary
Several major AI models hit version four on the same day, and this roundup from Theoretically Media covers them across video, language, image, and audio. The headline item is a first look at Kling 4, which has entered early access as the Flash model for annual Ultra subscribers, with the larger model planned for October.
The creator tests text-to-video, image-to-video with dialogue, and Omni reference generations, including a diner scene with an invented background character, a birthday scene that shows improved spatial consistency, and a noir detective scene. Output is currently 720p. The notes cover flat performances, an unprompted wardrobe change, and a technique of combining keyframes (up to 10 references) with Omni references for better acting.
The video also covers HeyGen Video, the company’s first general-purpose video model, a fine-tune of MiniMax H3 priced from 1 cent per second at 480p during a 50% launch sale through the end of October, and available via API and ComfyUI. Google’s announced Gemini 4 Argon is discussed, with impressive but unverified benchmarks and limited release. Updates from Ideogram 4.5 and ElevenLabs v4 round out the coverage.
📺 Source: Theoretically Media · Published October 01, 2026
🏷️ Format: Roundup
