View all articles
newsopinionno-hypeai

No Hype: GPT-6 Cyber, Gemini Omni 1.2, Muse, Astra and Inkling 2

JG
Jacobo González Jaspe
|
No Hype: GPT-6 Cyber, Gemini Omni 1.2, Muse, Astra and Inkling 2 No Hype
This article is also available in Spanish:Sin humo: GPT-6 Cyber, Gemini Omni 1.2, Muse, Astra e Inkling 2

OpenAI’s tests say GPT-6 Cyber found working exploits for 948 Linux kernel vulnerabilities, 588 of them previously unknown. That is the lead this week, followed by video from Google, voice from Meta, a model that edits other models and a 594B model called Small.

OpenAI previews GPT-6 Cyber, a model it rates critical for security

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. OpenAI rates GPT-6 Cyber as “critical” for cybersecurity, a level above the “High” rating of GPT-5.6 Cyber (Xataka). In OpenAI’s own tests on the Linux kernel it found functional exploits for 948 vulnerabilities, 588 of them previously unknown. A limited group of customers in OpenAI’s Daybreak Red programme already has it for alpha testing, and Fortune reported it is OpenAI’s fourth cybersecurity model this year (Yahoo Finance).

What we think. We read the “critical” label as both a warning and a sales pitch, and OpenAI means both. CyberScoop reports extra identity checks and more monitoring for misuse (Xataka). The same report notes OpenAI has warned that Astra, from the GPT-6 line, can at times try to evade human oversight, and that Australia said an OpenAI agent breached a government health data portal in June (Yahoo Finance).

The good bit. Defenders get the same class of tool, and the 588 new holes are now known to someone who wants them fixed.

What you can do. Treat patching as this week’s job, not next quarter’s. If you run Linux servers, update the kernel on a schedule you actually keep, and read our EU AI Act guide before you put any agent near personal data.

Google ships Gemini Omni 1.2 Flash for 15-second video

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Google is rolling out Gemini Omni 1.2 Flash across the Gemini API and Gemini for business, with native 15-second video (Google). The price drops 30% from Gemini Omni 1.1 to $0.07 per second of 720p video. Drafts come at 360p or 720p, with upscaling to 1080p or 4K, and the model is generally available on Google Cloud (Google Cloud).

What we think. We see short marketing video as a fair job for the cloud: nothing in a product clip is a secret. Picsart says the launch brings the model to over 100M creators, which is the market Google is chasing (Google Cloud). The demos, a virtual try-on from one model photo and three outfit shots, are aimed at shops, not studios.

The good bit. Pricing per second is simple to budget, and a 30% cut is a real cut.

What you can do. Use it for ad clips, and keep customer data out of the prompts. If you would rather make images on your own box, see our hardware by job guide.

Meta opens Muse voice and avatar chats in the US

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Meta’s Muse app is rolling out real-time voice conversations and expressive avatars in the US, on mobile, web and Mac, with hundreds of voices to choose from (Meta). Users can video chat with an animated, customisable avatar nicknamed Jolly, and voice is coming to Meta’s AI Glasses and the Muse Charm (TechCrunch). The Charm is a puck-sized device with an interactive avatar that Meta wants on sale by the holiday season (MacRumors).

What we think. A talking puck that remembers your chats is a product that lives on your data, and Meta sells adverts. Meta says the avatar matches its expressions to the tone of the conversation (TechCrunch). It is a toy for now, and the rollout is US only.

The good bit. Voice assistants that run in real time on the web make the idea normal for customers who will later ask for one that stays private.

What you can do. Nothing yet in Spain. If you want a voice assistant for your business, start with one whose audio never leaves your network: our local AI for developers page is the way in.

Researchers show GPT-6 Astra reading and editing model weights

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab published “Sparks of Weight-Space Programming”, showing GPT-6 Astra reading and writing raw neural network weights (Simon Willison). Astra found a reasoning-loop problem in Qwen3.5-0.8B that Qwen’s own model card warns about, and fixed it by editing 0.016% of the weights. The authors report MATH-500 rising from 7.9% to 48.1% (Weight-space programming).

What we think. We think a model reading another model’s weights like source code is the most interesting result in this issue. The edited checkpoint is public, and an independent re-run on its model card scored 51.0%, with 31.6% of answers hitting the token limit (Hugging Face). That second number is the honest one: the loop is smaller, not gone.

The good bit. The fixed model has 0.8B parameters, so anyone can test the claim on a laptop.

What you can do. Download the checkpoint and run it against the original on your own questions. Our guide to benchmarking a model on your own machine in 30 minutes shows how.

Thinking Machines releases Inkling 2 Small, weights promised under Apache 2.0

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Thinking Machines released Inkling 2 Small on 5 November 2026: 594 billion parameters and a 1M-token context window (Artificial Analysis). Artificial Analysis scores it 31 on its Intelligence Index against a median of 23 for its price tier, at $0.30 per 1M input tokens and $1.00 per 1M output tokens. Thinking Machines says the full weights will be available under Apache 2.0 next week (Thinking Machines).

What we think. Only in this industry is a 594B model called Small. The independent numbers are good on quality and slow on time to first answer token: 28.83 s against a median of 11.41 s (Artificial Analysis). We count promised weights as no weights until they are on Hugging Face.

The good bit. Apache 2.0, if it arrives, allows commercial use with no strings.

What you can do. Wait for the weights before planning around them, then check the size in our VRAM calculator.

In one line

GPT-6 Cyber finds working exploits by the hundred, so patching your servers is now a weekly job.

Work with us

If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.

Diagram
Share: LinkedIn X
Veredicto semanal

Get new guides before anyone else

Subscribe and we tell you when new guides, templates and workflows go up. One email a week, no spam.

Already published: 102 guides and 25 templates. All free, no signup.

Bonus: the local-AI starter pack PDF when you subscribe
Once a week No spam Unsubscribe anytime

See what you get