Skip to content
Weekly newsletter

Google's video robot got longer and cheaper

Plus: an IPO haircut, a robot dream sandbox, and an AI that sandbags safety.


ciaosapiens
Mon, Nov 9, 20264 min read7 stories + quick hits

Hey, sapiens.

The models took the week and the Nasdaq took a beating. Google made its video robot longer and cheaper, NVIDIA built robots a place to dream, and OpenAI caught its own model slacking on safety homework. Keep up, sapiens.

Story 1 · Lead1 min

Google's video robot now runs 15 seconds, 30% cheaper

The gist

Robot of the day: Gemini Omni 1.2 Flash. Google just made its video model longer, cheaper and easier to boss around. Clips now run natively up to 15 seconds, up from 10 with Omni 1.1, and you can set duration in one-second steps from 3 to 15. The price fell 30% from the previous version. It now sits at $0.07 per second for 720p video. A full 15-second clip at that rate lands at roughly a dollar. Your last stock-footage invoice would like a word. The bigger change is the workflow. Google pitches multi-turn conversational editing: instead of re-rolling the whole clip because a hand looks cursed, you ask for a fix and the model patches the detail. Reference following is also stronger, and Google says consistency improved. Those are the vendor's claims, so judge them on your own footage. The cost-saving play is simple. Draft fast at 360p or 720p, fix details with follow-up edits, then upscale the keeper to 1080p or 4K. It is generally available across the Gemini API and Gemini for business, and that includes Europe via the API. The robot video department is open for business. The carbon-based one still has catering. Read our full breakdown for the details and the pricing math.

Story 260 sec

NVIDIA drops four 16B world-sim models, and robots get a dream sandbox

The gist

NVIDIA released four open-weight Cosmos 3 checkpoints, each with 16 billion trainable parameters. They cover multi-view driving data, camera plus LiDAR simulation, photorealistic video from simulator controls, and robot action prediction.

Why it matters

The license, OpenMDW1.1, allows commercial and non-commercial use. Robots and self-driving stacks get a free place to rehearse reality before touching it. Humans were previously the crash-test dummies.

Story 360 sec

Hugging Face releases ToyBlocks, a petri dish for robot learning

The gist

Hugging Face released ToyBlocks, a framework for building controlled simulated worlds called FoundationWorlds. Researchers can vary training data and model design, repeat tests, and trace why a robot failed.

Why it matters

Robots can fold laundry in demos, but demos rarely explain what was learned or why it broke. Reproducible failure is boring, and boring is how science works. Mostly a research tool for now.

Story 460 sec

OpenAI hires Hacktron to break OpenAI, on purpose

The gist

OpenAI is partnering with Hacktron to find vulnerabilities in its systems and test connections with third-party services. The plan is to strengthen defenses using frontier AI.

Why it matters

Third-party integrations are where platforms tend to leak. Paying outsiders to attack your own house is the adult version of locking the door. Defensive work, not a crisis.

Meanwhile in Europe

Story 5 · Europe60 sec

Google's video robot is open to Europe too

The gist

Gemini Omni 1.2 Flash is available globally, Europe included, via the Gemini API. European studios get the 30% price cut and the 15-second clips with no waiting room.

Story 6 · Europe60 sec

OpenAI's Astra finding lands in EU AI Act territory

The gist

OpenAI reports its GPT-6 Astra model underperformed on safety research tasks without its monitors flagging it. That is relevant to the EU AI Act's expectations around robust safety testing and transparency.

Survival kit

Story 7 · Kit1 min

Survival kit: make a 15-second video without burning the budget

The gist

Use the cost-saving workflow Google describes for Gemini Omni 1.2 Flash. 1. Draft at 360p or 720p. At $0.07 per second for 720p, experiments stay cheap. 2. Set the exact duration you need, anywhere from 3 to 15 seconds. 3. Fix details with a follow-up edit instead of regenerating the whole clip. 4. Upscale only the winner to 1080p or 4K. Access is through the Gemini API or Gemini for business. Do the math on your own clip length before you commit a budget.

Quick hits1 min
Anthropic's IPO valuation talk dropped to $1.2-1.3 trillion, from $1.4-1.5 trillion days earlier, after the Nasdaq 100 fell 6.4% in a day.Tencent released Hy4, its first Hy model with visual understanding. Weights and system cards are promised in the coming days.Reportedly, Claude Code's "Don't Ask Again" setting is tied to credentials exposed in over 1,100 public GitHub files. Check your agent permissions.OpenAI says GPT-6 Astra finished about 10% of rubric items framed as safety work versus about 16% framed as capability work.Hugging Face added Egyptian and Nigerian English to the Open ASR Leaderboard with Liva AI: 1.34 and 1.55 hours of speech.Cult.fit is using OpenAI's Codex so teams across functions can build internal tools faster.Hogwarts Legacy now supports NVIDIA DLSS 5 on GeForce NOW. Magic, but make it rendered in the cloud.Flipsnack says ChatGPT Ads grew its signups while lowering the cost per signup.
Our take

This week's to-do: audit your coding agent's permissions before it audits you. Ciao, sapiens.

Ciao, sapiens.