Big AI Weekly: Gemini’s New Tricks, NotebookLM Upgrades, and Claude Mythos Warning

It was a big week in AI, with Google rolling out major Gemini and NotebookLM upgrades, Anthropic revealing a frighteningly powerful new model, and several AI agent platforms pushing forward. Here’s what actually matters and how it might affect the way you use AI day to day.

Gemini’s Biggest Upgrade Yet: Notebooks, Visual Answers, and Music

Google shipped several meaningful updates to Gemini that make it feel more like a real workspace and less like a single chat box.

NotebookLM Notebooks Inside Gemini

Paid Gemini users are starting to see a new “Notebooks” section in the left sidebar. This isn’t just a shortcut—it’s a full project-style organization system powered by NotebookLM.

Here’s what’s new:

• Notebook list in Gemini: All of your NotebookLM notebooks now appear directly inside Gemini under a dedicated Notebooks section.

• Per-project chat history: When you start a chat from inside a notebook, that conversation is stored inside that notebook, not in the general “All chats” list. Each notebook effectively becomes its own focused workspace.

• Bidirectional sync: Chats you have in Gemini inside a notebook also show up back in NotebookLM as sources. NotebookLM content flows into Gemini, and Gemini chats flow back into NotebookLM. That makes it much easier to keep research, notes, and AI conversations tied together in one place.

This is the project/folder system many Gemini users have been asking for, and it finally makes longer-term work in Gemini more manageable.

Interactive Visualizations for Complex Topics

Gemini can now turn certain questions into interactive visual explanations—especially for scientific or mathematical concepts.

How it works:

• You need to have Gemini Pro enabled.

• Ask a question that naturally benefits from a visual, such as “Explain Boyle’s law” or “Teach me how planes fly.”

• When Gemini decides a visualization makes sense, you’ll see a “Show me the visualization” button under the text answer.

• Clicking it generates an interactive graphic (which can take 10–60 seconds to load) that helps explain the concept.

It doesn’t appear for every query (for example, “What food should I bring on a hike?” won’t trigger it), and some visuals are more helpful than others. But it’s a promising first step toward more visual, intuitive AI answers—especially for students and self-learners.

Longer AI Music on the Free Plan

Gemini’s built-in music generation also got a small but welcome upgrade.

• Previously, free users were limited to 30-second tracks.

• Now, free users get a limited number of longer tracks per day when they choose the “Thinking” or “Pro” modes in the “Create music” flow.

• After you use up those longer generations, you can still create 30-second clips using the “Fast” mode.

The music quality isn’t meant to replace professional production, but it’s great for fun ideas, quick background loops, or sharing silly songs with friends and family.

Gemini for Home Speakers Expands Globally

Google is also rolling out Gemini for home speakers to 16 new countries with support for 7 additional languages. If you have a compatible Google speaker, you’ll be able to use Gemini as your home assistant in more regions, with more natural language support.

NotebookLM Gets Smarter for Studying and Learning

Beyond the deep integration with Gemini, NotebookLM picked up two quality-of-life upgrades that make it more useful as a learning tool.

Smarter Quiz Reports and Targeted Practice

NotebookLM’s quizzes and flashcards now generate a detailed report after you finish a session. The report includes:

• How well you performed overall

• Which topics were covered

• Specific areas where you struggled

From there, you can click on a weak topic and instantly generate a new quiz focused just on that area. It turns NotebookLM into a more adaptive study partner that helps you double down on what you don’t know yet.

Better Controls for Audio Overviews

Audio overviews—NotebookLM’s spoken summaries of your material—now have more playback controls:

• Adjust speed from 0.5x to 2x

• Skip forward or backward in 10-second jumps

This makes it much easier to treat NotebookLM like a personalized podcast for your notes, research, or study materials.

New Google AI App: Offline-First Voice Dictation

Google quietly released a new free app called Google AI Edge Eloquence, focused on smart voice dictation.

What it does:

• You tap a button, speak naturally, and the app transcribes and cleans up what you say.

• It can remove hesitations and corrections (for example, if you say “one week… I mean two weeks,” it will keep only “two weeks” in the final text).

The standout feature is that it can run completely offline on-device. You can toggle offline mode and still get decent transcription and cleanup without sending audio to the cloud—good for privacy and when you don’t have a connection.

However, the experience is not yet seamless:

• There’s no keyboard integration, so you can’t use it directly inside your messaging apps.

• You have to switch into the app, record, stop, copy, and paste the text manually.

• There’s no support for system-wide shortcuts or action buttons yet.

Because of that, it’s unlikely to replace more integrated tools like Whisper-based workflows right now, but it’s an interesting glimpse at private, offline AI utilities coming to phones.

Anthropic’s Claude Updates: Open Claude Cutoff and the Mythos Shock

On the Anthropic side, there were two important developments: a policy change that affects power users, and a new model that’s powerful enough to raise serious security questions.

Claude Subscriptions No Longer Power Open Claude

Anthropic announced that Claude consumer subscriptions (like the $20/month or $100/month plans) can no longer be used to power Open Claude, an open-source always-on AI agent framework.

Previously, some users connected their personal Claude subscriptions to Open Claude to drive autonomous agents. Anthropic now requires that these setups use the official Claude API instead, which is priced for developers and businesses and is significantly more expensive.

If you’re not using Open Claude, this change doesn’t affect you. But for advanced users and tinkerers, it’s a major shift in how affordable it is to run powerful Claude-based agents at home.

Claude Mythos: The Most Powerful (and Restricted) Model Yet

The bigger news is Claude Mythos, Anthropic’s new ultra-capable model that’s currently too powerful to release broadly.

From what’s been shared so far:

• Mythos appears to be the most capable AI model in the world right now, especially for tasks like code analysis and security research.

• Anthropic has reportedly used Mythos to uncover serious software vulnerabilities, including bugs that have existed for nearly three decades.

Because of how easily it can find and potentially exploit security flaws, Anthropic is not making Mythos publicly available yet. Instead, they’ve launched Project Glasswing and are giving tightly controlled access to:

• Selected government agencies

• A small group of large companies (reportedly including Apple and possibly Google)

The goal is to let these organizations find and patch vulnerabilities before models at this capability level become widely accessible. It’s a cautious approach that has sparked debate about safety vs. openness. If you want a deeper dive into what’s known so far, check out this breakdown of what we actually know about Claude Mythos and the broader discussion in whether Mythos is a powerful cyber defense tool or a massive security risk.

For everyday users, the practical takeaway is simple: keep your devices updated. As models like Mythos expose more long-standing vulnerabilities, OS, browser, and app updates this year will be more important than ever for staying secure.

Other Notable AI Updates: Perplexity, ChatGPT, and GenSpark

Beyond Google and Anthropic, several other AI players shipped meaningful updates.

Perplexity Computer Now Connects to Your Finances

Perplexity Computer—one of the most advanced AI agents available—can now connect to your financial accounts via Plaid.

Once connected, it can:

• Pull in data from your bank accounts, credit cards, and loans

• Analyze your spending in detail

• Help you build custom budgeting tools

• Visualize your net worth alongside your investment portfolio

Plenty of budgeting apps already exist, but this is essentially that idea powered by a highly capable AI agent that can reason over your data, answer questions, and build tailored workflows.

New $100/Month ChatGPT Pro Tier

OpenAI introduced a new $100/month ChatGPT Pro plan, filling the large gap between the standard $20/month plan and the $200/month tier.

This mid-tier option is aimed at users who have outgrown the basic plan—often heavy professionals or small teams—but don’t need or can’t justify the full enterprise-level pricing. Expect higher limits and more generous access to advanced models compared to the $20 plan.

GenSpark: Easier Full-Desktop AI Control

GenSpark is an AI agent platform that has been gaining traction, especially among people who found Open Claude too complex or risky to run.

Its latest update adds:

• Full desktop control: The agent can open apps, work inside Microsoft Office (Docs, Sheets, Slides equivalents), and generally control your computer like a virtual assistant.

• WhatsApp remote access: You can control your desktop agent remotely just by messaging it from your phone.

• Heartbeat checks: The agent can periodically wake up to perform tasks, such as checking your email every few minutes.

Combined, these features make GenSpark feel like a more approachable, user-friendly alternative to highly technical agent setups, while still offering serious automation power.

Overall, this week showed both sides of the AI story: more helpful tools and quality-of-life upgrades for everyday users, and behind the scenes, models like Mythos pushing the limits of what’s possible—and what’s safe.