# Neues KI-Modell Gemini 4 Carbon schlägt Claude & ChatGPT, Duet Agent & Genie 4 | KI-News der Woche

Channel: Digitale Profis
Video: https://www.youtube.com/watch?v=kMTTSVcZmKM
Duration: 14 min
Language: English
Words: 2313
Transcript page: https://viewrankai.com/tools/youtube-transcript/kMTTSVcZmKM

---

[0:00] Claude is getting an assistant that reaches out to you on its own. Google is already following up with its next top model, Gemini 4 Carbon, showing interactive AI worlds with sound, and new Flux models are really speeding up image and video generation. As always, we’ve summarized all the important AI news from the past week for you. You can find the links to our sources, more information, and our newsletter in the

[0:22] video description below, as usual. Let’s get started. Claude "Duet" keeps working in the background. We’ll start with Anthropic and Claude "Thought." The idea behind it. The new assistant summarizes your upcoming tasks from your connected apps in the morning and notifies you during the day if something needs your attention. You can also hand over larger tasks, such as preparing a presentation or finding outstanding invoices. The trick: Duet works on its own virtual computer, complete with a browser. Your own laptop doesn't have to stay open for it to run. However, before the assistant does anything critical, like sending a message or triggering a payment, it shows you its plan and waits for your

[1:01] approval. Under the hood, it is powered by Claude Opus 5. Currently, the preview is only starting via a waitlist for Pro and Max subscribers in the US. Here in Germany, we will have to be a little more patient. But there is one catch you should keep in mind now. Duet shares the usage limit with your normal Claude account, and the regular

[1:20] background checks count towards it too. An assistant that independently looks for new tasks will eat into your quota, even when you aren't actively interacting with it. Google introduces Gemini 4 Carbon. Things are happening fast at Google right now. Gemini 4 Argon was released just two weeks ago, and now Google is already following up with the next top model, Gemini 4 Carbon. Carbon takes the number one spot in the updated Artificial Analysis Intelligence Index and is primarily optimized for complex programming, IT security, and extremely long agent tasks. According to Google, the model ran for days at a time in internal tests to work through dead ends and

[2:01] find and fix deep-seated bugs in large code projects. At the same time, Carbon is said to make half as many hallucinations compared to Argon and, in tests, did not let a single attack through regarding protection against prompt injections. Because the model has become so much better at finding and exploiting security vulnerabilities , Google is proceeding with caution at launch. Starting this week, selected IT security researchers and partners will

[2:27] get access via Google's "Fairwind." Developers, companies, and regular users like us should follow soon at the same introductory API price as Argon, i.e., $ per million input tokens and $, before the price later increases to $ 4 and $ 20. Until Carbon is open to everyone, you can use Gemini 4 Argo directly in the meantime. It is now available to all Pro and Ultra subscribers in the Gemini app and via the API. Fittingly, the Gemini app has received a new task mode, where the AI continues working in the background even when you leave your computer. And in the Mac app, you can also use the Gravity programming environment

[3:10] directly. GPT 6.1 Terra: OpenAI is also catching up on computer control and is releasing GPT 6.1 Terra. The model is specifically designed to understand screen content and operate applications directly, e.g., to transfer invoice data into accounting software or fill out forms. Terra is also expected to show improvements in writing compared to its predecessor, GPT 5.6 Terra. The model is being introduced in Jet GPT

[3:37] Work, in Codex, and via the API. According to OpenAI, it cannot be selected in the standard chat. The price is particularly exciting for developers. At $ 1 per million input tokens and $ 5 for output, the tokens cost exactly half as much as the previous Terra. And with GPT 6.1 SO, OpenAI simultaneously reports progress in programming and fewer factual errors in its own tests. This price drop is especially noticeable for AI agents, as many model calls are often required in the background for a single task. Terra is now expected to make exactly these processes faster and cheaper. However, for particularly difficult programming and research tasks, OpenAI still

[4:18] recommends the larger models, Soul or Astra. Genie 4 creates interactive worlds with sound. Google DeepMind provides a rather impressive look ahead with Genie 4. It allows for the creation of virtual environments that you can move through live and change while in motion. You can spontaneously change the weather, place a new character in the scene, or adjust the camera angle. The biggest leap this time: Genie 4 now generates the appropriate sound along with it. The environment comes with matching noises, and characters are also supposed to

[4:49] react to you and speak in real-time. Let's take a look at that with sound.

[5:15] Create a scene, move the camera, directo and the In practice, you could use this during film pre-production, for example, to play through different camera angles and sequences before shooting on set. Game developers could also try out an idea as a walkable environment first. Google, however, points out the limitations itself. During longer runs, details can change, and the physics also remains somewhat error-prone. So far, only selected testers are allowed access. Next, Ultra subscribers in Project Genie and early API partners are planned. PHP vulnerabilities discovered with AI affect website operators. Now a message that is important for everyone who operates their own website. On the 21st . In October, OpenAI and the PHP team

[6:14] publicly released two critical security vulnerabilities in PHP along with the corresponding updates. PHP runs directly on the server and forms the foundation for WordPress, among other things. Under certain conditions, attackers could execute their own code via the vulnerabilities without needing to log in beforehand. It is exciting how these bugs were discovered. The research team Kalif used GPT 5.6 Soul to track down one of these vulnerabilities. The security firm WH had also independently discovered and reported both vulnerabilities. Before the release, all parties involved worked behind the scenes with the PHP developers and major infrastructure providers on fixes and protective measures. Our advice: if you run a website, check briefly with your hoster

[6:56] to see if your PHP version has already received the update from the 21st. of October. We have also covered this topic in detail in a separate video. You can find the link to that video and the accompanying article with practical tips in the video description. Flux 3 Max accelerates video generation. Now moving on to video generation. There is now Flux 3 Max from Fal, developed in

[7:17] collaboration with Black Forest Labs. The model is based on Flux 3, was further trained by Fal, and specifically optimized for speed. According to published measurements, a 10-second video is created in approximately 12 seconds. Anyone who has generated multiple variants of a scene knows what a difference that makes when experimenting. Normally, you wait for minutes between individual attempts. Flux 3 Max is available directly via Fal, for creating videos from a text description or a starting image, among other features. Regarding quality, the provider points to the top rankings in the video arena by Artificial Analysis. How well the material is suited for your own editing , of course, always depends on the

[7:56] subject. If the results are right, however, the short waiting time helps enormously with correcting movement and testing different camera settings. Flux 3 Image Small for local image generation. Black Forest Labs also presents Flux 3 Image Small. The compact image model not only creates new subjects, but also edits existing images and can combine up to ten reference images. By drawing areas, you can specify exactly where an element should be placed or which part of an image should be changed. Super practical if a product photo is largely correct and you only want to exchange a single object. Via the API, there is an additional feature that preserves all

[8:33] unchanged image areas exactly. Especially exciting for local setups, the model weights are available for free download. According to Black Forest Labs, a downscaled 8-bit version runs on Nvidia graphics cards with 8 GB of VRAM, for instance, directly via ComfyUI. This makes custom image workflows really interesting, even without large server hardware. A base version for personal customization and a speed-optimized variant are being offered. The Flux Community License applies to its use. If you want to use the model for client projects, you should check the license terms beforehand. If you just want to try it out, you can find a demo on Hugging Face in addition to the API. ChatGPT

[9:12] Slides creates collaboratively editable presentations. Back to the more boring office tasks. OpenAI is introducing ChatGPT Slides. You simply start with a chat history, a file, or an outline and have a complete presentation built from it. You can also provide an existing sample presentation as a design guide. After that, you can edit texts, images, and diagrams either directly on the

[9:36] slide or have them changed via prompts. Especially practical for teams: multiple people can work on the same presentation in a Space simultaneously and leave comments. On the web, you can then export the slides as PowerPoint, as a PDF, or transfer them directly to Google Slides. The feature is being rolled out to eligible Pro, Business, and Enterprise accounts with Space access. Instead of tediously copying text back and forth, you work directly as a team on the draft that ChatGPT just created for content review. In the end, however, some manual work is still required for the moment. Voxa 3 for transcripts and voice agents. There is

[10:10] finally some news from Mistral again. With Voxa 3, the company has updated its speech models. The family includes models that transcribe recordings and distinguish between different speakers. A real-time variant as well as a separate model for speech output. A practical detail for recognition in around 50 languages. You can provide terms in advance that are likely to occur in the conversation. This helps enormously when many product names or special technical terms are used in a task. For this, Mistral provides an SDK that allows developers to build their own voice agents. This connects listening and speaking directly with a language model and custom software tools. A telephone assistant could thus

[10:47] check for an available appointment and prepare the booking during the call. Mistral offers both an EU-hosted API and operation on your own infrastructure for this, so companies can decide for themselves where recordings are processed. The new models are launching as public research for now. Meta releases Mu-Spark Open. Meta has released the model weights for Mu-Spark Open. The large multimodal model is intended especially for programming tasks and longer agent workflows. In addition to text, it processes images, audio, and video as input and generates text responses. It can be used either via various API providers or completely on your own infrastructure. However, you need a significant amount of RAM for your own

[11:29] operation. In total, Mu-Spark Open comes with 816 billion parameters. Even if only a portion of these are active at each step, the entire model weights must fit into memory first. Meta itself is demonstrating the model on, among other things, a setup of Mac Studio systems with M5 Ultra chips. For companies and specialized teams, this opens up the possibility of running and customizing such a model on their own hardware. Commercial use is subject to Meta's own Llama Model License. Kimi k 1.5 Pro with new coding and 3D functions. There is also a new model from Moonshot. Kimi k 1.5 Pro has been available since the 23rd of October in

[12:09] Kimi Code, in the other Kimi products, and via the API. The provider promises noticeable improvements in longer programming tasks and especially in working with 3D content. With Kimi 3D Make, models can be viewed and edited directly in the browser and downloaded for 3D printing, for example. In the demos, Kimi independently operates 3D tools and builds entire scenes from them. Additionally, there is a high-speed variant that, according to Moonshot, responds two to three times

[12:36] faster at the same quality level. Context length and model variant depend on the respective subscription plan. While you can already use the new models in Kimi, the full model weights are scheduled to be available for download by November 4th at the latest. November at the latest. Runway builds creative apps from a description. Finally, a word on Runway. With the new App Studio, you can easily build your own small creative tools using text descriptions. You describe what the application should do, and Runway creates the appropriate interface for it. In the background, Runway's existing image, video, and audio models work combined with your own files,

[13:08] brand kits, and finished workflows. Initial examples include a tool for product images and an application with which camera positions can be specified in three-dimensional space. For teams, this solves a well-known everyday problem. Previously, only one individual often knew the exact prompts and settings. With App Studio, a marketing department can prepare such a tool once, and everyone else on the team just inserts a new product photo later, without having to search for the image style specifications every single time. Runway hosts the finished application directly for paid accounts, so you can use it in your team or share

[13:41] it with other Runway users via a link. And that concludes our AI update for this week. As always, I would appreciate likes for the video and subscriptions to our channel. My name is Timothy Michener. See you next time. Bye.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt