This discussion analyzes leaked details concerning Google DeepMind's forthcoming AI releases: Gemini 3, Nano Banana 2, and a new agent prototype. Gemini 3, developed through costly "hero runs" and iterative checkpoints, exhibits unparalleled one-shot coding prowess, exemplified by functional Mac OS and Windows replicas generated in the browser and complex SVG rendering. An early, superior checkpoint, dubbed ECPT, demonstrated mastery over International Math Olympiad-level problems. However, the anticipated Gemini 3 Pro preview (rumored release: November 18th) shows a slight, likely efficiency-driven, reduction in coding ability compared to ECPT, favoring a more balanced model profile. Nano Banana 2 is concurrently positioned as a major leap in image generation and editing quality, and Google is also testing a web-based task automation agent.
Summary of Leaked Google DeepMind AI Products
0:03 Gemini 3 Capabilities: Gemini 3 is deemed the best model yet, demonstrating exceptional competency and speed. Leaks show complex one-shot coding, such as generating functional, interactive operating system replicas (Mac OS and Windows) directly in the browser using clever JavaScript tricks to simulate functionality.
1:49 Leak Source: Hero Runs and Checkpoints: The leaks stem from Google's extensive and costly "hero runs"—full-scale training efforts combining all successful experiments and concept upgrades. The model is released in periodic "checkpoints" to external users for crucial feedback, explaining the variability and early access.
3:27 Accessing Early Versions: Users are gaining access by repeatedly prompting in AI Studio until a "side-by-side comparison" appears, indicating an answer provided by a variant other than the current Gemini 2.5 Pro (potentially a Gemini 3 Pro branch or the leaner Flash version).
4:23 The ECPT Checkpoint (Peak Performance): The most powerful leaked version is referred to as "ECPT," showcasing extreme one-shot abilities, including creative coding (e.g., generating 3D Voxal art of a "Pelican on a bike"), interactive scientific visualizations (exoplanet core), and functional interactive clones (Minecraft).
6:51 Mathematical Aptitude: Gemini 3 allegedly solved Somara’s 554th problem (an International Math Olympiad-level difficulty previously solved only by GPT-5 Pro), accomplishing the feat in 3 minutes (compared to 15 minutes for GPT-5 Pro).
7:37 Skeptical Gaming Example: An output generating a complex, functional, playable game complete with menus, damage animation, and a soundtrack in a single shot was shared, though the speaker expresses high skepticism regarding this specific claim.
8:40 Gemini 3 Pro Preview and Downgrade: ECPT is reportedly not the final version. The Gemini 3 Pro preview is anticipated to launch around November 18th (based on Google Vertex console deprecation entries), but users report a slight downgrade in coding ability compared to ECPT.
10:30 Efficiency Trade-off: The perceived downgrade is likely intentional, indicating Google chose a more robust and balanced model, potentially optimizing for lower token usage and computational efficiency over the peak, though potentially less reliable, performance of the ECPT branch.
11:36 Nano Banana 2 Upgrade: The image generator Nano Banana 2, a significant upgrade over Nano Banana 1, was accidentally exposed briefly. It boasts better quality, instruction following, 2K native output, 4K upscaling, and transparency support, with a likely release around mid-November.
13:07 DeepMind Agent Prototype: DeepMind is testing a new agent prototype designed for web task automation (e.g., web surfing on the user's behalf) and complex design commands within tools like Google Stitch, currently limited to a research preview stage.
13:54 Market Response: The pace of these leaks suggests an aggressive end-of-year competition, with OpenAI expected to respond quickly, possibly with something like GPT 5.1.