# Gemini Omni 1.2 Flash, Cosmos 3 checkpoints, Atlas live, PDD Kit

> Google's 15-second video model, four open Cosmos 3 checkpoints, World Labs' real-time Atlas demo and Microsoft's proof-driven coding kit.

Published 5 Nov 2026. Coverage: 5 Nov 2026 to 5 Nov 2026. By Ali Asif.

## Items

1. **Introducing Gemini Omni 1.2 Flash** (Google, Frontier): Video generation and editing now reaches 15 seconds per clip, up from 10, at 30% lower cost than Omni 1.1, or $0.07 per second at 720p. [Primary source](https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemini-omni-1-2-flash)
2. **New NVIDIA Cosmos 3 Open Models for Interactive World Simulation** (NVIDIA, Open weights): Four open-weight checkpoints post-trained from Cosmos 3 Nano, covering multi-view driving data, streaming camera and LiDAR, simulator-to-video and two-arm robot prediction. [Primary source](https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation)
3. **Frameboy: Atlas in Real-time** (World Labs, Image, video, world): A research preview that streams frames from the Atlas world model to a browser at up to 15 fps while you steer the camera. [Primary source](https://www.worldlabs.ai/blog/atlas-realtime)
4. **Formally verifying agent-written full-stack web apps with PDD Kit** (Microsoft Research, Agents and coding): A research prototype that turns developer intent into a Verus specification, proves agent-written Rust against it and labels each requirement as proof or test evidence. [Primary source](https://www.microsoft.com/en-us/research/blog/formally-verifying-agent-written-full-stack-web-apps-with-pdd-kit)

Four primary pages I could open and read today: a video model, a set of open world-simulation checkpoints, a real-time world demo and a verification kit for agent-written code. Everything below comes from those pages, not from my own testing.

## Gemini Omni 1.2 Flash from Google

Google calls Omni 1.2 Flash generally available. Clips now run up to 15 seconds, up from 10 in Omni 1.1, and duration can be set in one-second steps from 3 to 15 seconds. Price drops 30% against Omni 1.1, which Google puts at $0.07 per second for 720p video.

The workflow advice is practical. Google suggests drafting at 360p or 720p, fixing details with follow-up edits, then upscaling to 1080p or 4K. Inputs can include character, product or style images, motion clips up to three seconds and audio cues. It is available through the Gemini API in Google AI Studio, in Google Cloud with Provisioned Throughput, in Flow and in Google Vids, and it is rolling out in the Gemini app for Google AI Ultra subscribers. Generated videos carry a SynthID watermark.

The quality claims are qualitative. Google reports less drift in characters and products from first frame to last, fewer unwanted jump cuts in camera moves and clearer lip-sync with distinct voices for multiple speakers. It also says fast sports action, rapid script changes and complex physics are still in progress, and I appreciate that it said so. There are no benchmark numbers on the page.

## Cosmos 3 open checkpoints from NVIDIA

NVIDIA added four open-weight checkpoints, all post-trained from Cosmos 3 Nano. Transfer-Auto generates multi-view driving data with up to 11 camera views and joint RGB and LiDAR output. Sim-Auto is a causally distilled streaming variant that produces synchronized camera video and LiDAR in small chunks. Sim-Transfer turns simulator controls such as depth, segmentation and edges, plus a text prompt, into RGB video. Sim-Bimanual predicts the next observations from actions for YAM and AgiBot two-arm robots.

The article says 16B trainable parameters per model, while the Hugging Face cards list 15B. The license is OpenMDW1.1, which permits commercial and non-commercial use under its terms. The streaming recipe has four stages, ending in consistency distillation at four denoising steps per chunk and self-forcing. The vLLM-Omni setup was tested on one GB200.

## Frameboy from World Labs

Two months after introducing Atlas, an autoregressive world model, World Labs released Frameboy as a research preview. It streams frames from cloud GPUs to your browser at up to 15 fps while you control the camera, using an optimized Atlas variant. The same mode is wired into Viewfinder, a web app for building scenes frame by frame, which is in closed beta. World Labs says plainly that Frameboy needs a high-end cloud GPU server and is not cheap next to local rendering.

## PDD Kit from Microsoft Research

PDD Kit is a research prototype for proof-driven development. It converts developer intent into a Verus specification, has a coding agent write Rust and the proof, and falls back to tests where a proof cannot reach. Each requirement is labeled as proof or evidence, and a mutation check plants one bug per ensures clause to catch clauses that constrain nothing. The case study is a roughly 5,000-line Rust app where the team verified the login logic. The page does not link a PDD Kit repository, and it verifies only Rust today.

## What I'd watch next

Whether Google publishes numbers for Omni 1.2 Flash, and whether PDD Kit gets a public repo.

## My take

Omni 1.2 Flash is the biggest launch, but PDD Kit is the one I would read twice, because it asks the developer to review intent instead of code. I read all four pages and ran none of the tools, so every result here is the publisher's own.

## Sources

- https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemini-omni-1-2-flash
- https://huggingface.co/blog/nvidia/cosmos-interactive-world-simulation
- https://www.worldlabs.ai/blog/atlas-realtime
- https://www.microsoft.com/en-us/research/blog/formally-verifying-agent-written-full-stack-web-apps-with-pdd-kit

Canonical: https://heyaliasif.com/ai/2026-11-05-gemini-omni-1-2-flash-cosmos-atlas-pdd-kit