# Новости ИИ: Новые модели ChatGPT и Google (14 обновлений, которые стоит знать)

> Полный текст доступен на английском

**Published:** 6 ноября 2026 г.
**Author:** Matt Wolfe
**Source:** YouTube — Matt Wolfe
**Language:** en-US
**HTML:** https://aiqazaqstan.com/ru/articles/ai-news-new-chatgpt-google-models-14-updates-worth-knowing

Узнайте, что могут автоматизировать для вас Virtual Teammates от Optimizely по адресу http://optimizely.com/mattwolfe На этой неделе в новостях ИИ: GPT-Image 2.6 от OpenAI, Google Gemini Omni 1.2 Flash и новейшие инструменты ИИ для создания изображений, генерации видео и выполнения задач. Смотрите демо…

Welcome to the weekly deep dive of everything you need to know that happened in the world of AI this week. Let's start with the new image model out of OpenAI. This week they introduced images 2.6. It's available in both chat GPT and in the API and apparently it's faster and more professional visuals. We can see in their little demo here it does go, you know, almost twice as fast. But the big sort of magic trick of this new model is that it can make animated GIFs for you or animated GIFs, whatever you prefer. So if I jump into chat GPT here, I could give it a prompt like make an animated GIF of a monkey roller skating. And like 20 seconds later, we got our monkey roller skating here. Let's see if I click on it. I do have an option to edit it. So, it actually looks like it tried to put it in San Diego, which is probably based on the memory that it knows I'm in San Diego. But what happens if I go like make him wearing a baseball hat and make him roller skate in the street? So, I'll give those two edit prompts. Send them both in. And another 30 seconds or so later, he's now got a baseball hat and he's in the street instead of the sidewalk. So, it seems like the editing works just fine even on animated GIFs. It does feel a bit faster. I mean, I'm surprised cuz it's making a GIF, which means it's probably making, I don't know, eight different images or something and then stitching them all together. And it did it all in about 30 seconds or less. So, let me test how long it takes to make just a regular image. Created an image of the San Diego skyline at sunset with sailboats on the water. So, I'm going to submit it and I'm going to start my timer here. Ready? submit. So, this is what it generated. It looks pretty solid, and it did it in 20 seconds. 20.4 seconds. So, not too bad. Definitely a speed improvement from what we used to get out of the previous image model. But, I'm going to go ahead and move along. We did get a lot of updates this week, but I would say they're all kind of smaller updates, so I don't need to go super deep into any of them. So, it's almost going to be like an entire rapid fire episode, but I will even do a rapid fire later. But moving along, we got another image model this week that I think is worth mentioning real quick. This company, Pruna AI, released a model called PH3 image. Now, they actually took the Miniax H3 model, which is a state-of-the-art video model, and turned it into an efficient image model, but not any image model, an actual open image model. So, this is one that theoretically you could download and run on your own device if you have enough RAM to do that or, you know, run it on some sort of cloud provider, but you can fine-tune it. You can alter it. You can do whatever you want with it cuz it's open weight. We can see some example images that it generated here. The left is Quinn image 2.1. The right is this new PH3 image model. Here's like a realistic image of an old fisherman. Here's an anime image. And we can see the weights that you would download are about 55 gigabytes, but they do have some quantitized models that are a bit smaller that you could run as well. Here's some more examples of images it generated. Here's even more. Now, like I mentioned, there's a few variations of it. You've got the 16 bit, the 8 bit, and then the 8bit VAE down here. And if we look at the memory requirements to generate an image at 8 megapixels, it requires about 56 GB of VRAM at the 16 bit and 21 GB at the 4bit. Meaning that this 4bit quantized model, you should be able to run on some consumer grade GPUs. Like this would run on a 5090 or a 4090. Now, this one you're going to need a little bit beefier hardware. So, if you're going to run the full version, it looks like you would need, you know, like a Mac Ultra M3 or M5 that has that, you know, 128 GB of unified memory or up. This would also run on something like a DGX Spark, but you need something pretty decent to run the full version of this. Now, you can actually use this model yourself. Here's some instructions on how to use it on your own device. However, they do have it over on Hugging Face here. This is where you would download the weight. So, if you did want to download it onto your own system, the entire thing is almost 100 GB. But we can see here they have a hugging face space. And a hugging face space means that they went and they set it up on hugging faces GPUs to allow people to use it and test it directly inside of hugging face. So, you can actually use this one for free right now online. It might be a little slower depending on how many people are trying to use it at the same time, but if I gave it a prompt like the San Diego skyline at sunset with sailboats on the water, it will generate right here in hugging face. Now, the image quality definitely nowhere near what the chat GPT image quality is. Like that looks nothing like the San Diego skyline at all, but it is openweight and you can fine-tune it. And if you have a decent enough device, you can run it locally if you want. Another cool update comes from Optimizely. I talked about their new virtual teammates in a previous video, but I've been able to play around with their preconfigured SEO and AI search analyst teammate, and it's pretty cool. I just connected it to my futuretools.io site, set up the virtual teammate, and then I asked it, "Audit future tools for SEO and AI search visibility, and tell me what to fix first." It can research keywords, find content gaps, and check how ready my site is to show up in AI answers. That's the AEO part of it that everyone is focusing on these days. It stands for answer engine optimization because so many people are getting their recommendations from AI instead of traditional search. Now, this virtual teammate helps me with all of that and it can continue to monitor the site and proactively collaborate with me on prioritized action items. So, I've got a running list of what needs attention without manually digging through dashboards and figuring it out all on my own. If you want to check it out or any of their other virtual teammates, head to the link in the description box. And thanks so much to Optimizely for supporting my channel and sponsoring this portion of today's video. We also got an update out of Google this week. They have a new video model, their new Gemini Omni 1.2 flash. I think the biggest update about Omni 1.2 is that it can generate up to 15-second clips. I believe the previous model only generated up to 10 seconds. So, we get five extra seconds on the clips that we're generating. The price is also lower. So, they dropped the price 30% from Omni 1.1 and now it cost about 7 cents per second of 720p video. Some of the claims of this new model is better reference adherence. So, we can see the reference image of this woman and then three different outfits. And on the right, we can see how well it maintains both the woman's image and the outfits that she's trying on. It claims more consistent results as well. So, we can see the reference image they gave it. And then on the right, uh, we have Omni 1.1 Flash and 1.2 Flash and 1.2 Flash definitely looks a lot closer to the reference image. When it comes to editing, it's apparently much better at editing. So, if we look at this video here, we can see the original video. And then over on the right, we can see four different variations with four different backgrounds, but the people at the forefront all look exactly the same as they did in the original video, only changing the background. It also claims more natural speech. So, let's take a listen to this. Excuse me. That orchid is copying my color palette. I demand royalties. >> I think it just wants to match your tail, Cosmo. >> Technically, it bloomed here last Tuesday. Man, don't beef with the plants. Yeah, not bad. More controllable cinematic looks and of course longer clips. So up to 15 seconds. Now it's available in the Gemini API as well as inside of Flow and Google Vids. But the cool thing is Google is letting anybody use it even on a free plan. You can actually generate up to three videos a month on a free plan. We can see here starting today, anyone with a Google account or Google Workspace subscription can generate up to three 15-second videos per month in Vids at no cost. So, it looks like you need to be using the Vids platform to generate the three videos. But, if you head over to vidsnew, this is where you can use that model for free up to three times a month. So, over here in the bottom right, we can see create from scratch. So, let's go ahead and add an avatar. I'll use this one of my face. It looks like I didn't know I was taking a picture there. I looked pissed off or something, but whatever. Let's use it. Add an ingredient here. Now, I'll add this blue shirt. And then for the prompt, I'll say me is wearing. And then I'll put this image there and saying subscribe to Matt Wolf on YouTube while a bear juggles chainsaws in the background. And let's set it to 15 seconds. Let's just go to the max. Now, that was actually pretty quick. I'd say it took maybe about a minute, maybe slightly less than a minute. >> Subscribe to Matt Wolf on YouTube. >> Not too bad. So, pretty impressive. 15 seconds. It was quick to generate and it was free because it was one of my three per month. Now, at 7 per second, 15 seconds, it would have cost me about $15 if it wasn't one of the free ones. Since we're talking about video models, might as well mention this one as well. Pika Labs late last week announced Pika H3 Flex, which is an openweight video model. It can generate up to 4K. It's faster than real-time generation, and it can also produce 15-second clips with soundtracks in as little as 7 seconds. That is crazy. And on Design Arena, we can see it pretty much beats out like everything. All the other openweight models, it beats out. Now, they did also make this one available on hugging face, so you can download the model weights here. It's 220 GB, so it is quite large. And I do believe if you want to run it locally on your own system, you do need a pretty solid system. Let's see what it says here. So, for their stated results of generating videos in 5 seconds, they did do it on an H100 with 80 GB of VRAM. So that's not really going to be like a consumer GPU. So again, this is one where you're probably going to need to run it in a cloud unless you have like an M3 or M5 Ultra with a huge amount of unified memory or you're using something like a DGX Spark, you're going to need solid VRAM to be able to run this. But again, it is openweight. It can be fine-tuned. You theoretically can run it locally. And similar to the earlier openweight model we were looking at, they do have a hugging face space. So, they did set up a space where you can go and generate directly on hugging faces GPUs and test it out for free. This is the default prompt that was here. Let's generate and see what it does. So, it took about 30 seconds, but here's what it generated. I mean, it looks pretty good. It's sort of a boring prompt, but it does look good. Just for old times sake, let's do Will Smith eating spaghetti. And let's go 15 seconds. Is it going to block us? Ah, it blocked us. Oh, no. It didn't block us because of Will Smith, though. It blocked us because of the GPU duration. So, I think if I bring it back down, let's try like 8 seconds. It's like Grandpa Will Smith. But, it took about 30 seconds to generate also. So, it seemed like bumping it from 5 to 8 seconds didn't cause it to take a whole bunch extra time to generate. So, again, you can use this one for free right now over on Hugging Face instead of the Hugging Face space. It will likely be slower if there's a lot of people all trying to use it at the same time, though. And if you have a solid enough computer, you could download this and run it locally on your own system. Moving right along, we got some updates out of Meta and their Muse model. You can now have voice chats directly with MetaMuse. In fact, not only can you have voice chats, you can have video calls with Metam Muse. I'm trying really hard not to say Metamucil. So, here I am in my Metamuse account. Let's create a new side chat. And let's click up in the top right. You can see we've now got a little call button. So, let's call Gary. >> Hey, Matt. Happy Thursday afternoon. What's new? >> Not much. You don't sound like your name would be Gary. >> Give me just a second. I've switched to a mid-pitched, warm, velvety, middle-aged male voice that should feel a lot more like a Gary. How does this new one sound? >> Sounds great. Um, I have a question for you. What can you help me with right now given everything you have access to? >> With everything you've got connected, I can help in a few key ways, Matt. I can handle your Gmail and calendar. Keep your content studio ideas flowing for Instagram and that push to 100K, plus research, drafts, and scheduling. Want me to check today's inbox, your calendar, or pull your latest Instagram ideas to start? So, from today's refresh, you've got a same day hands-on with that new Gemini video model, the swoozy AI animation debate, that Steam AI games stat with the Valve and Epic angle, and a quick safety check for Clawude Code users, plus a couple evergreen carousels. Want me to pull the full script for the Gemini one first? >> I think we get the idea, though. We can just now talk to our Muse instead of having to type to our Muse. And it can do everything that Muse has been able to do already. And we also have a little like avatar that we can see while we're talking to, but you know, I don't really know the point of having the avatar, but it's there. Being able to use voice does give much more of like a talking to an assistant vibe, though, versus typing. I feel like I can say, "Hey, go check my emails and calendar and tell me what I should do next." And things like that. I'm not going to because it will probably share sensitive information, but again, you get the idea. Moving on, if you're a user of Grockbot, they just rolled out a new feature where Grockbot can now make phone calls on your behalf. Now, in my opinion, Grockbot is super underrated. I've mentioned it in past videos. A lot of people refuse to use things like Grockbot because they're not fans of Elon, but it actually is a really, really good agent. I've been a fan of it. I have not tested this phone call feature. And so just for fun, I want to see what it sounds like if it was to call me. So I'm going to tell it to call me on the phone and act like I'm a business. And let's see what happens. All right. So I've got Grockbot open here. And I did ask, is it true that you can make phone calls for me now? And it says, yep, I can. So I'm going to tell it to call me, but I'm going to tell it it's like a business number. I'll just say call this number. Obviously, I'm blurring out my number. and asked them about their hours today. It's for a comic book store called Matt's Comics. So, let's see what happens. So, it called me out and was like, "We looked it up and that's not a real business phone number." And it says, "If this is actually your own phone or someone you know and you're testing how calls work, let me know who it belongs to." Yes, it's my phone. I'm testing this out to see what calls sound like. Okay, so it's asking me to accept the call. So, go ahead and call. And a call's coming through. Let's see what happens. Hello. >> Hi, I'm here. >> You're very, very quiet. Let me put you on speaker phone. One second. Okay. Uh, what did you say? >> No worries. I said I'm Leora, Matt's virtual assistant. Just doing a quick test call. Can you hear me better now? >> I can hear you just fine now. How are you today? >> Great. I'm doing well. Thanks for asking. How about you? >> I'm doing great. >> Glad to hear it. Were you wanting to try asking about today's hours or anything else? >> I just wanted to know what you sound like when you call a store. Can you pretend that I'm a store and you're calling to ask about the hours? >> Sure thing. Hey there. I'm a virtual assistant calling for Matt Wolf on a recorded line to ask about your hours. Do you happen to know what time you open today? >> Uh, we're open right now. >> Perfect. That's all I needed. Want to give anything else a shot? >> I think that's it. Thank you. Yeah, I mean it doesn't sound that bad. Pretty cool that we can have our agents make calls for us. I do have a feeling like 5050 chance the second the person answers and here's it's a robot they hang up. I know if like I answer a phone call and I can tell it's like a rooc call I immediately just hang up and don't waste my time. And I have a feeling like stores and stuff are probably going to do that a lot too until this becomes a little more common. But we'll see. It doesn't sound too bad. All right. So I know I've kind of run through everything quick. It's sort of been a rapid fire type episode, but there are a few more things I do want to tell you about, but I just don't think they need a ton of time dedicated to them. So, this is like the actual rapid fire section, but it's going to be rapid firewai did actually roll out another model this week in GPT6 Cyber, but as you can probably guess, nobody gets access to it except for cyber security professionals. And of course, it beats every single benchmark on, you know, cyber security type stuff. OpenAI also made an update this week to chat GPT and work and codecs. Now the usage is sort of blended together for all of it. Usage allowances will be shared across chatgpt and codeex. The openw weight model, Kimmy K 3.1 Pro, which is basically the most stateofthe-art open weight model available now, actually released their weights this week. Now, I actually already talked about Kimmy K 3.1 Pro a couple weeks ago, and we dove a little deeper into it. The TLDDR of it is it is pretty much the best model available that is open weight, falling just barely below Claude Fable 5.1, which if you remember just a few months ago was like the best model anybody's ever seen, ever. And Kimmy K 3.1, an openw weight model, is like right there with it basically. Again, the model was made available like a week and a half ago, but the weights for this model, for the open weight version, are now available on hugging face. They are 1.56 tab. It's a 2.8 trillion parameter model, so you're probably not going to be running it on any home GPU right now. Anthropic rolled out a new feature in Claude this week as well, where you can actually connect your bank account into Claude. It's rolling out this week on Pro, Max, and Ultra plans. And it uses Plaid to connect your Claude account straight into whatever bank you're using. And you can do things like ask where did my money go this month. It'll give you a breakdown. You can ask like what's left before payday. They do not train on any of your financial information. And it is read only. So they can't like do anything to your bank account other than read information. And they only get access to like the last four digits of your account number. So they can't even see account numbers. I haven't connected my bank account to Claude yet. I'm a little hesitant. Maybe I'll try it in the future. But even if I did, I probably wouldn't show it off on video because I don't want to share that information with you. No offense to you. We have a new AI app from a person I never expected we'd get one from. Steve Ahoki partnered with somebody to create an app called Selendis. And it's an AI video editing app. It's in early access right now. I haven't gotten access yet, so I haven't been able to test it. But apparently, it's a personal AI video editor that learns each creator's taste from decisions made in the editor, the past video catalog, and reference videos that you site. Now, I put out a video earlier this week that shows my AI editing workflow. It looks like this, but I use Claude to edit my videos, and I have like a really, really killer workflow. So, watch that video if you haven't already. Netflix announced that they're adding even more AI into their platform where they will show recommendations through a new AI discover tab. I don't quite know how this is like super different from their existing categories, but it'll have things like something light, family movie night, quick watches, and then it'll use AI to sort of organize stuff based on your mood for what you want to watch. But I also feel like that's kind of how Netflix categories work already, so I don't know if we'll notice a big difference. Okay, these next two stories are really fascinating to me. We keep on hearing all these use cases around these AI agents ordering Door Dash for us or ordering us an Uber or we just pull out our voice mode and say, "Hey, uh, I'm feeling like getting some tacos today. Uh, have some tacos sent to my house and then, you know, in 2 hours make sure an Uber arrives so I get to the airport on time." And we see examples like that all the time of these are the killer use cases for agents. Well, it turns out nobody's using those features. According to Door Dash, they said that they've only received several hundred orders from AI agents. That was in the third quarter earnings call. And they do on average 3 million deliveries daily. So only hundreds out of 3 million daily deliveries are using AI agents. And that's 3 million daily. And this only seen a several hundred orders is not daily. That's like they've seen several hundred orders total, but they're doing 3 million orders daily. So like in the grand scheme of things, people are not using agents for that. Uber said something similar. This article on Yahoo Tech, almost no one uses AI agents to book their Uber. This also came out in their third quarter earnings call. Out of 320 million trips, only 500 Uber Eats orders were placed by an AI agent. So again, the theme kind of is across the board. People aren't using their agents to do things like order food and order rides and things like that. It's just not quite happening yet. And finally, Google's Project Suncatcher actually saw some results. On October 1st, they launched their prototype satellite, Project Suncatcher, which is basically GPUs or their TPUs up in space. And finally, this week, they reached their hello world milestone, the first ever AI response from a data center in orbit. Now, it was just a single response and it's just a test and it's just a prototype. So, it's not like at scale yet or anything, but they actually have a working satellite with GPUs or TPUs on it sending data back to Earth and AI models actually running up in orbit. That's pretty damn cool. And that's pretty much all the AI news this week. It has been a bit slower of a week, which is why I didn't spend a ton of time on any one story. And I think there's a few reasons for that. We had the elections going on in the US and a lot of these big companies were sort of waiting to see what was going to happen with these elections. We've been seeing the stock market kind of tank probably as a result of the election results. We also have Anthropic right on the verge of their IPO. I talked about that quite a bit in last week's video. And so Anthropic's probably waiting until after the IPO to announce anything big, right? Because if they IPO and then shortly after the IPO make some sort of big announcement, well, that has some potential to move the stock price a bit, right? I also think OpenAI is sort of watching and waiting to see what happens with the anthropic IPO. That's probably going to inform how they approach their own IPO. But also, OpenAI has this tendency to wait for a company to make a big release and then make their own big release right around that time to sort of overshadow that company. So, I think that's why it's been a couple slower weeks in a row. Election uncertainty, anthropics IPO, trouble with the stock market, all of that is contributing to a slower release period right now. But again, that's what I got for you. Every single week, I keep up with the news all day, every day. I'm sort of in it down the rabbit hole, drinking from the fire hose, keeping up with everything. So, at the end of the week, I can make one video that recaps it all for you, so that anybody can stay totally looped in on the entire world of AI by watching just one video a week. That's my goal with these videos. If you like that kind of thing and you want to stay looped in with just one video a week, make sure you like this video and subscribe to this channel. And I will do my best to make sure you're on top of everything. I also like to make tutorials and workflow videos and show you how to use AI in between my Friday videos. So again, if that's something that interests you, maybe this channel is a good one to subscribe to. But thank you so much for hanging out with me and nerding out with me today. I really, really appreciate you spending time with me. And hopefully you feel a lot more looped in. See you in the next one. Bye-bye.

---

*Материал агрегирован из первоисточника: [YouTube — Matt Wolfe](https://www.youtube.com/watch?v=IrAHINf0M6E). Открыть оригинал.*

**Теги:** #OpenAI #YouTube

---

*AI Qazaqstan — non-commercial informational portal about AI in Kazakhstan. Machine translations and summaries are AI-generated.*