Major model releases: OpenAI, Google, Alibaba, MiniMax, Microsoft, Krea, and multimodal competition
Key Questions
Which new image and video models were recently released?
Releases include Krea 2 (SOTA open-weights 12B), ChatGPT Image 2, Gemma 4 12B, MiniMax M3, Alibaba Qwen3.7-Plus, Microsoft MAI-Image-2.5, Ideogram 4.0, Reve 2.0, Grok Imagine Video, and Google's Nano Banana 2 Lite and Gemini Omni Flash.
What benchmarks are emerging for evaluating AI video models?
Physical plausibility evaluation (PQSG) is becoming a key benchmark, alongside pricing comparisons across 144 models on platforms like Replicate, Fal.ai, and DeepInfra where the lowest cost reaches $0.0002 per image for SDXL Turbo.
How is AI video investment shifting?
Investment is moving from raw models to studios, exemplified by the A24-DeepMind deal, with focus on tools like Google NotebookLM auto-generating 60-second vertical video summaries using Nano Banana 2 Lite.
What capabilities does Gemini Omni Flash offer?
It supports conversational video editing with multi-turn interactions, physics-aware generation, and provenance features, representing a UX shift from prompt engineering to iterative dialogue with a 3-5 edits per session sweet spot.
How are companies like Nvidia applying generative vision AI?
Nvidia launched vision AI agent blueprints for industrial use, including synthetic data for rare defects, video search, and Omniverse simulation, achieving 95% AP with 8 real images plus synthetic data at partners like Corning and Foxconn.
What 3D and world model developments were announced?
Tripo AI raised $150M for 3D and world models, while ByteDance Seedream 4.5 offers unified generation and editing at $0.04 per image, and Alibaba introduced the Qwen-Image-Agent framework.
How does Google NotebookLM integrate new image models?
It now auto-generates 60-second vertical video summaries powered by Nano Banana 2 Lite, enabling fast, low-cost content creation with approximately 3-second generation times at around $0.03 per image.
What trends are visible in multimodal AI competition?
Models like Seedance 2.5 (30s 4K clips with audio) and Kling 3.0 compete on native audio and editing, while open-weights options such as Krea 2 challenge closed systems on performance and accessibility.
New model comparisons: Krea 2 SOTA open-weights 12B image model; ChatGPT Image 2; Gemma 4 12B encoder-free; MiniMax M3; Alibaba Qwen3.7-Plus; Microsoft MAI-Image-2.5; Ideogram 4.0; Reve 2.0 4K; Grok Imagine Video. Google Project Genie, Seedance 2.5 (30s, 4K), Kling 3.0 with native audio, Google DiffusionGemma, Apple Image Playground photorealistic. Physical plausibility evaluation (PQSG) emerging as key benchmark. ByteDance Seedream 4.5 (unified gen/editing, $0.04/image). Alibaba Qwen-Image-Agent agentic framework. AI video investment shifting from raw models to studios (A24-DeepMind deal). Google Nano Banana 2 Lite (fast/cheap image gen, ~3s, ~$0.03/image) and Gemini Omni Flash (interactive video editing). Nvidia launches vision AI agent blueprints for industrial generative vision (synthetic data for rare defects, video search, Omniverse simulation; Corning 95% AP with 8 real images + synthetic, Foxconn yield improvement). Pricing comparison across 144 models on Replicate, Fal.ai, DeepInfra — lowest $0.0002 for SDXL Turbo. Fable 5 testimonial shows agentic screen recording editing capability. Tripo AI raises $150M for 3D and world models, signaling strong capital validation. Stitch Fix expands AI image generation for e-commerce personalization. Gemini Omni Flash guide: conversational video editing, multi-turn, physics-aware, provenance features. New: Google NotebookLM auto-generates 60-second vertical video summaries using Nano Banana 2 Lite. New: Grok Imagine guide (2026) with thread continuity tip for brand consistency. Nano Banana 2 Lite conversational interface: genuine UX shift from prompt engineering to iterative dialogue. Practical testing shows 3-5 edits per session sweet spot. Key for product tooling. New: Seedance 2.5 maturation highlighted in 'The Quiet Shift' article, showing 30s clips, multimodal inputs, audio-aware generation.