What's hot in AI: 4 to 6 Nov

Wednesday to Friday, 4 to 6 November 2026. Each edition stands alone.

Midweek our browser tabs fill with render queues and model drop-downs, tools that make things. By Friday the conversation shifts to what those same tools might break, or what new systems exist to break things before the wrong people do. This week the split felt almost deliberate. Makers on one side, defenders on the other, and the rest of us in the middle trying to remember which hat we are meant to wear.

OpenAI's Images 2.6 landed on Wednesday with quiet language that still hides real pace. Fifty percent faster generation, GIF support at last, and two new API models called Flare and Sunburst under the GPT-Image-2.6 umbrella. The company says people create roughly four billion images a week across ChatGPT, Codex and the API. I have not counted them. What I did notice is the SVG and markup editing. You can hand it a vector file or some raw code and nudge it visually. For a small agency, that means fewer round trips between design exports and developer handoff. GIFs matter less for polished client decks and more for quick tests of ideas that may not survive the week.

Tencent released Hy4 the same day. It is the first Hy model with both vision and computer use, trained inside real 3D environments: Blender, CAD, Unreal, Unity and Godot. The WorkBuddy benchmark score hit 83.18 with a 96.15 percent pass rate, and the hallucination rate dropped from 12.95 percent to 7.2 percent. Pricing stayed flat with the preview at roughly ninety US cents per million input tokens and two dollars seventy per million output. Weights are coming, which for a certain kind of weekend tinkerer is the real news. I have not run Hy4 yet. My Blender skills peaked years ago and I am not eager to confirm that in public.

Mr. Mummu turns up in these posts now and then, usually with the wrong tool and a proud explanation. When he heard that new helpers work best with "full permissions", he ticked every box on a trial assistant his nephew had mentioned, then went to make tea. The tool sorted his downloads into folders named after emotions. He was furious about "Regret" but grudgingly impressed by "Nostalgia." His Urgent notebook, mostly shopping lists, gained a fresh line under onions: "FULL PERMISSIONS = FULL PROBLEMS." Later he shrugged and said the sentence that mattered: "Decide the fence before you open the gate." Wrong tool first. Right truth by accident. Thirty seconds of ticking boxes, an afternoon spent undoing them, and somehow he still landed on the point.

Thursday also brought Google Gemini Omni 1.2 Flash into general availability. Native video clips extend to fifteen seconds now, up from ten, with multi-turn conversational editing that lets you talk through changes rather than restart a prompt from scratch. Pricing dropped thirty percent to seven US cents per second for 720p, and you can draft at 360p or 720p then upscale rather than burning budget on a first cut. SynthID watermarking is baked in, which matters for provenance even if most viewers will never check. This one is reachable for me, and short product loops without the usual production pipeline look like the right first test.

Then OpenAI dropped GPT-6 Cyber on approved Daybreak Red customers, and the tone of the week shifted. This is not a creative tool. ExploitGym scores jumped from 32.1 percent for GPT-5.6 Cyber to 48.5 percent. It stayed in scope one hundred percent of the time when given deployment prompts, which sounds dry until you realise what "scope" means here. OpenAI's own research turned up 948 Linux privilege-escalation exploits and 81 novel high-severity Firefox vulnerabilities. The system is classified Critical under their Preparedness Framework. Access is gated and I do not have it. For small teams, the capability is real and the access is tightly restricted. Do not treat this as something to play with on a slow afternoon.

Google also made Gemini API Computer Use generally available on Wednesday without breaking changes, so the news moved quietly. I mention it because computer-use features shipped from more than one provider in the same window, each with different fences.

Most of this column traces company announcements. I have not personally run GPT-6 Cyber or Hy4. Omni 1.2 Flash and Images 2.6 I can reach and likely will. The numbers above come from OpenAI's Introducing Images 2.6 post, Tencent's Hy4 announcement, Google's Gemini Omni 1.2 Flash blog, and OpenAI's GPT-6 Cyber post. The Computer Use availability was noted in AI Compass coverage of Google AI Studio and the Gemini API docs. Nothing here is professional, financial or security advice. It is opinion and education from one desk.

This morning's Vibe News episode on my YouTube channel covers a related week in ads, agents and creator tools.

Whether you are generating a GIF, rendering a scene in Unreal, or probing software for holes, the same question applies. What is this tool allowed to touch? Makers and defenders both need a fence. The creative tools this week got faster and cheaper, which means more people will reach for them with less forethought. The defensive tools got sharper and more gated. Decide the leash before you plug the next tool in. Your downloads folder does not need a folder called Regret.