We finally got Google’s new flagship model, Gemini 4 Argon this week. This model was announced last month but was only available to select testers until a couple days ago. It’s now widely available. Today we’ll explore the model, how it stacks up against the competition, and look at what you can create with it. Google shared a number of interesting examples from building racing games to creating a 3D modeling app that outputs files for printing.

What is Gemini 4 Argon?

Gemini 4 Argon is Google’s new frontier model. It’s built for long, complex tasks that involve coding, business, cybersecurity, and science. Google explained that Argon currently outperforms both GPT-6 Astra and Anthropic’s Fable 5.1 across numerous benchmarks. One of the biggest changes facilitating this performance is the amount Argon can produce for each response. It is able to write up to one million tokens in a single response, the highest of any frontier model. Essentially, this means that Argon can work through much larger problems in one go instead of taking several steps. Google also explained that the quality of its answers has improved, with more direct answers and less hedging.

Argon was initially announced on September 30th but was only given to select beta testers and cybersecurity experts for additional testing. The trusted security experts were given a version of the model without the usual cyber safeguards so they could use its full abilities to find and patch holes in their systems.

What people have built with Argon

Google shared some great demos this week and some users have already surpassed them. Here are a few examples that stood out to me:

• Desk Racing Game: Argon used photos of an actual desk and turned the objects on it into a playable racing game. It plays surprisingly well with working physics, sound effects, and good graphics. You can play and remix this game for free!

• 3D Wizard Chess Game: This was another fun example of what Argon can build. This game was created with Google’s coding agent Antigravity and used Python to create all the 3D elements. You can play it here!
• Interactive 3D experience: In this example, Argon used only code to create a 3D interactive recreation. The level of detail and customization is remarkable.

• Interactive Heritage Site: Argon was able to turn LiDAR scans and research papers into an engaging 3D experience. You can click on each of the buildings to read snippets from the papers while exploring the site.

Who can use it

If you pay for a Google AI Pro or Ultra subscription, then you can already use Gemini 4 Argon in the Gemini app. If you are an Ultra user, you can also access the model via AI Studio and in Antigravity. Developers can also use Argon through the Gemini API at “introductory pricing” that is reduced by 50% until the end of November. Unfortunately, free users can’t access Argon yet and no plans have been announced.

My Thoughts

Gemini 4 Argon has been a long time coming. We were promised Gemini 3.5 Pro months ago at Google I/O and then never got it. The delay makes sense when we look at the other models that have been released since then. GPT-6 Astra, Fable 5.1 and even Opus 5.5 are incredibly advanced. So in this environment, it’s understandable that Google would hold off on releasing a model if it couldn’t meet or surpass the competition. But late or not, I’m glad we’re finally getting a new high-end model from Google. As someone who loves vibe coding in AI Studio, I can now use a state-of-the-art model in my apps.

Google has also followed the precedent set by Anthropic by granting early access to cybersecurity specialists. By putting Argon in the hands of trusted beta users, they can better ensure the safety of the model before public releases. This is a smart move, especially as models get powerful enough that some of their abilities need to be limited or blocked before a public release.

Gemini 4 Argon is Here was originally published in Learn Prompting on Medium, where people are continuing the conversation by highlighting and responding to this story.

Last week, OpenAI released their long-awaited GPT-6 Astra to a select group of trusted testers. This week, we got our hands on this new model and could see the incredible tasks Astra was able to complete. We’ll explore in detail what makes GPT-6 different, some of the wildest things it created, and how it compares to Claude and Gemini.

What makes Astra different: it uses the software for you

In the past, chatbots could only explain how to complete a task that required an external tool like Unity or Premiere. They were limited to writing out explanations and couldn’t actually finish a task for you. So if you asked it to build a 3D model, it would explain what you needed to do to create it in Blender. But GPT-6 Astra is different. It can use these tools for you and actually build models in Blender or edit videos in Premiere. It can open the menus, drag the objects, write its own code, and even check its own work to ensure accuracy. So now instead of a step-by-step guide, you’ll get a finished product.

OpenAI refers to this ability as computer use and it’s what makes Astra so special. Aside from this powerful ability, GPT-6 also has some of the highest scores on browsing, coding, cybersecurity, and professional benchmarks. Currently GPT-6 is only available through the paid plans (Plus, Pro, Business, Enterprise) with no release date for free users.

What people are building with it

The last week has seen some awesome demos of what Astra can do. One of the most impressive examples comes from Ammaar Reshi, who leads product and design for Google AI Studio. He used Astra to port Windows versions of iconic AAA games like Skyrim, Age of Empires 2, and more running on his iPad. Astra created unique touch controls for each game and was able to hit 60 fps in some cases. While this alone is a fun and impressive example, it gets even crazier. At one point, Ammaar told Astra “when I wake up I want Batman: Arkham City running on my iPad”, and while he slept, the model used screen mirroring by itself to watch its own progress and make fixes.

The examples only get more interesting from there and go well beyond gaming. One user created a 3D map of a human ankle to help him understand a medical diagnosis. He can look at the skeleton or just the muscles and see a detailed overlay that includes clickable labels.

How it stacks up against the others

The obvious question whenever a new model gets released is, how does this compare to the models I already use? GPT-6’s main selling point is computer use, where it operates websites and software to finish tasks for you. But how does it compare when looking at standard reasoning tests? Astra can compete with other frontier models, but doesn’t beat them across the board. Fable 5.1, which we talked about last week, still outperforms it on at least one major reasoning benchmark. Currently Google doesn’t really have a model that’s capable of beating GPT-6 but 3.8 Flash isn’t really meant to.

The main tradeoff of GPT-6 is price. Astra can be expensive to run and is comparable to Anthropic’s top tier models. So the real deciding factor for choosing a model should be the task. If you’re doing research or asking a question, using Fable might make more sense. If you’re completing a task and want AI to do the heavy lifting, then Astra is probably the best choice.

My Thoughts

This is the first time I’ve ever felt that the idea of AI “doing work for you” has seemed real. Having a model that can open a tool it wasn’t specifically trained on and figure it out in real-time is such a huge step forward. GPT-6 doesn’t feel like just another release but instead like a real milestone in AI. Despite its incredible capabilities, I still wouldn’t call it AGI. I would say that this model feels like the closest thing we’ve gotten but still falls short of that lofty title. What I keep coming back to is the shift this marks in AI. We’re watching AI shift from finding and explaining things to just doing it for you. I’ve never used Blender before and now I have an AI model that can do it for me. While I still don’t see Astra fully replacing experts, it is enabling people to create projects that weren’t feasible two weeks ago.

GPT-6 Is Here. See What People Made With It was originally published in Learn Prompting on Medium, where people are continuing the conversation by highlighting and responding to this story.

This week I wanted to talk about one of my favorite AI tools: NotebookLM. This standalone, Gemini-powered tool is an AI research assistant that instantly becomes an expert on the sources you upload. This means that instead of pulling from unreliable internet sources or outdated training data like a traditional chatbot, it focuses solely on the documents you choose to share. This week we’ll explore its core differences, how best to use it, and look at some of NotebookLM’s best features.

What makes NotebookLM different

The thing that sets NotebookLM apart is that every response is source grounded. You get to upload your files and everything it tells you comes directly from those sources. In fact, NotebookLM won’t answer questions that are unrelated or not mentioned in your sources. Every answer also includes a citation that goes directly to your source so you can double check information or continue researching the answer on your own. As you’d expect, this effectively eliminates the potential for hallucinations and makes it the perfect tool for breaking down massive amounts of data or content for written reports and research papers.

NotebookLM supports multiple source types so you can upload just about any file type. PDFs, Google Docs, Word files, spreadsheets, images, websites, and even YouTube videos. Another difference is that instead of just the standard question and response loop, you can create unique outputs like Audio and Video Overviews.

The features worth your time

NotebookLM can generate a ton of helpful outputs, but several of them are worth covering here.

Audio Overviews are the feature that first made NotebookLM popular. They can turn your sources into a podcast style output where two hosts talk through your files and help make the material more engaging and digestible. This feature also includes an Interactive Mode that allows you to speak directly to the hosts. By asking questions and guiding the conversation, you can better understand large amounts of information.

Video Overviews are the visual version. Similarly to Audio Overviews, this feature creates an engaging output that consists of a narrated slideshow with unique visuals that pull information straight from your documents. These are great for visual learners.

The Quiz and Flashcard features fit perfectly for learning. By uploading the text or lesson, you can get a custom, interactive learning experience. The Quiz feature doesn’t just create a static multiple-choice quiz, but will instead give you hints and explainers for each question.

Deep Research brings the ability to find relevant sources for your notebook. This feature is focused on finding and suggesting sources that will complement or contradict your existing materials. This is a great resource for when you just start a new notebook and may not have tons of sources yet.

The Notes feature is simple but highly effective. You can save anything NotebookLM creates or write your own notes, and keep them inside your notebook. This ensures that as you work in your notebook, you can pull out the most important information and store it separately as opposed to losing it to your never-ending chat history.

These aren’t the only tools that NotebookLM offers, Mind Maps, Reports, Infographics, and Slide decks can also be created.

A few real ways to use it

The easiest way to understand NotebookLM is to jump right in. You can drop in a stack of research papers or a dozen technical reports and quickly ask questions or pull out the most important information. Alternatively, you can use NotebookLM as a resource for learning. By finding a dense academic paper and uploading it to a notebook, you can then create flashcards and quizzes for each chapter. And when there is a long piece of content that you want to absorb but you don’t have the time to sit down and read it, you can listen on the go using Audio Overviews. You can also just use NotebookLM as a more traditional chatbot and take advantage of the grounded responses to effectively work with your own data or content. At the end of the day, NotebookLM is a flexible tool that can be utilized for a number of different tasks.

My Thoughts

NotebookLM stands out to me because it’s clear about what it’s meant to do. So many AI tools overpromise and underdeliver, but NotebookLM is different. At its core, this tool is meant to give grounded responses based only on the sources you share. These responses come in a variety of forms but are trustworthy and evidence based. If you go in expecting a ChatGPT clone then you’ll be disappointed, but if you want to work with 50 files in one notebook to pull accurate, evidence-based answers then NotebookLM is the perfect tool. Of course, the only way to know for sure is to create your first notebook!

Note: This post was originally shared in the Learn Prompting newsletter. You can join here to get weekly issues just like this!

Google’s Most Useful AI Tool Isn’t Gemini was originally published in Learn Prompting on Medium, where people are continuing the conversation by highlighting and responding to this story.

When Nano Banana was first released last year, it was hailed as one of the best image generators to date. Now, Google is once again vying for the top spot with its new model, Nano Banana 2. This isn’t just a minor release; it’s a major leap forward. Nano Banana 2 was designed to be smarter, faster, and produce higher quality visuals than any other model on the market. This week we’ll break down the core changes and see this new model in action.

What’s New?

Nano Banana 2 brings sweeping new upgrades for all users. These are a mix of new features as well as some that were previously locked behind the paywall of Nano Banana Pro.

Instruction Following and Text Rendering:

Google’s new model is better at following your precise instructions. This means that you can give complex prompts with far more detail and Nano Banana 2 will create the visuals you need. When we combine this with better text generation, you get a model that can be used for any number of creative tasks. Anything from creating marketing materials and social media assets to stunning comics like the one below.

Character Consistency:

A trend in recent image models has been an improvement in character and object consistency between related images. Google has continued to build on this concept with Nano Banana 2. They shared that you can successfully maintain 14 subjects in a single workflow.

Smarter Model:

This new model can pull relevant information from its extensive knowledge base. This effectively gives an image model the same “intelligence” as the Gemini text models. Nano Banana 2 can also search the web to find and correctly render specific objects or subjects. We can truly see the power of Nano Banana 2 when we combine knowledge retrieval and image generation; creating visually appealing and accurate diagrams, infographics, and other data visualizations.

Visual Fidelity:

Nano Banana 2 brings improved image quality across the board. Specifically images now have “vibrant lighting, richer textures and sharper details” and up to 4k resolution. This means that whether you’re creating visuals for work or personal projects, you’ll get far better results.

How to Access Nano Banana 2

Google’s new image model is rolling out across all platforms. Here’s how to access:

• AI Mode: You can access the new model in Search via AI Mode.
• Gemini App: Nano Banana 2 is the new default image model for all users.
• API: You can access the new model via the API and Google AI Studio.
• Google Labs: Nano Banana 2 is the new default image model in video generator Flow. It is currently available to all Flow users.

My Thoughts:

I’ve always found image models to be a great resource. Quickly creating unique visuals or diagrams means that I can focus on other elements (like writing) without getting bogged down by my lack of artistic ability. That’s why I’m always excited to explore the latest AI image generation models. Nano Banana 2 did not disappoint. The combination of quality, consistency, and speed sets a new baseline for all future AI image models. But don’t just take my word for it.

Head to Gemini and start creating to see for yourself what you’re most excited about.

This post was originally posted on the Learn Prompting newsletter. Sign up here for our weekly email newsletter!

Nano Banana 2 Is Here — And It’s The Best Image Model Yet was originally published in Learn Prompting on Medium, where people are continuing the conversation by highlighting and responding to this story.