In partnership with
AIHOY and happy weekend, AInauts!
Maybe you missed some of last week's AI news, tools, and hacks. Or maybe you've only just joined us. No worries!
Here's our recap of the week's highlights:
We'll finish with the essential Quick News, so you can catch up on the latest developments in one place. Ready? Let's go!
Your AI prompts are leaving out 80% of what you're thinking.
When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning (constraints, edge cases, examples, tone) and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.
Used by teams at Microsoft, Amazon, and Notion. Try Wispr Flow free on Mac, Windows, iPhone, and Android.
Half an hour in the car, a few spoken instructions, and ChatGPT is working on an invoice spreadsheet, an email draft, and the next community events. That's exactly what we tried.
With plugins, custom app connections, and persistent agents, chat is becoming a place where we get work done rather than just talk about it. But the important boundary is elsewhere: customer records still belong in the CRM, invoices in the accounting system, and finished results in files we control.
Privacy-first email. Built for real protection.
End-to-end encrypted, ad-free, and open-source. Proton Mail protects your inbox with zero data tracking.
Hit the usage limit again, even though the task wasn't that big? Often, the most expensive model is doing the work at the highest reasoning setting, rereading half a novel with every step. We've put together six ways to change that, from tightly scoped tasks and suitable models to a clean start when you switch topics.
Haiku 5.5 also makes smaller tasks cheaper. Switching everything over blindly would still be the wrong way to save. Pick one recurring job, compare the result and consumption, and keep the stronger model where mistakes are expensive. A short answer alone doesn't prove low usage, either. The thinking beforehand costs tokens too.
Four hours of Elon Musk is a commitment. We watched the new documentary, made without his active involvement, which uses an AI avatar to reproduce some of his statements. It paints a very different picture from the familiar success stories about rockets, electric cars, and the next big thing.
It leaves us with an uncomfortable question: how much trust does a provider deserve just because its product is good? Grok can be useful without us handing its owner all our data and access. The same question applies to OpenAI and Anthropic. Choosing a model and trusting a company are two different decisions.
Gemini 4 Argon and Muse Wonder are here. So we put them up against GPT-6 Astra and Claude Opus 5.5: five everyday tasks in German, three runs per model. Checking invoices, finding contradictions in contracts, spotting numerical errors, seeing through invented sources, and writing usable copy. The 60 runs cost $8.72 in total.
Our small practical test doesn't replace an independent benchmark. But it shows how close the models have become on many normal tasks, and that even a top model can miss something.
You turn off model training and assume your chat is now private? It's not that simple. A recent study measured actual network traffic from nine consumer chatbots and found transmissions to third parties. That's a different question from what a model remembers.
We separate storage, training, tracking, and sharing through links or integrations. Only then can you sensibly decide whether a text belongs in an ordinary chat, needs a protected work environment, or shouldn't be uploaded at all. Business plans were not examined in this study.
Want to compare answers from three models without buying three more subscriptions? That's exactly what we use OpenRouter for. One account gives you pay-as-you-go access to different models, including side-by-side comparisons on the web.
You can also connect OpenRouter to your existing assistant and get a second opinion there. Just remember: those calls aren't covered by your existing ChatGPT or Claude subscription.
AI News Quickie: The Industry Highlights
AI never sleeps!
This week, assistants sound more natural, reach deeper into your browser, and get new models to choose from. And Brussels is enforcing the AI Act for the first time. Who controls increasingly independent agents is no longer a theoretical question.
💬 OpenAI: Better Conversations, More Sites, and a Farewell
ChatGPT Slides now builds presentations from conversations, files, or an outline, with your team working directly on the same slides. Rollout starts with Pro, Business, and Enterprise accounts with Space; on the web, export to PowerPoint, PDF, and Google Slides.
The ChatGPT Chrome extension can now read, compare, and work in your open tabs, even without the desktop app.
GPT-Live-1.1 and its smaller sibling Quip are rolling out in ChatGPT Voice, with more natural accents and new voices, including German. Choose them under Settings → Voice. API access is coming too.
For ChatGPT Sites, Plus and Pro lose the fixed cap on the number of sites. Usage and storage limits still apply.
GPT-5.5 has officially retired from ChatGPT, Work, and Codex, but remains in the API. If a saved workflow still requests it, switch that workflow now.
For developers: with Prompt-to-Voice, describe your voice agent's voice in words instead of choosing from a list. Available to all API customers, with voice creation free. The Decisions integration is new too: the agent can decide where a request belongs while the conversation is still happening.
Among OpenAI's small and medium-sized EU customers, weekly active users of Codex and ChatGPT Work have grown more than eightfold since May.
An OpenAI engineer was arrested by the FBI after allegedly hiding a homemade listening device in a conference room used by executives. OpenAI says there is no evidence that models, code, or customer data were taken.
🔎 Google: Argon Arrives, Scams Get Easier to Spot
The new Gemini 4 Argon launches worldwide, in the Gemini app for Pro and Ultra subscribers and through the paid API. Introductory pricing of $2 per million input tokens and $10 per million output tokens runs through November 30, then doubles. Our everyday test is above.
Google is also giving 10,000 researchers six months of free access to Antigravity's highest usage tier. Applications are open, with admission happening in stages.
Gemini Notebook now turns your materials into 60-second explainer videos, interactive study overviews, and new quiz formats such as fill-in-the-blank and short-answer questions.
Android apps you build in AI Studio can reach your phone privately through Google Play, without a developer account. In the same update, Ask Play compares apps based on their differences and reviews.
Suspicious text message on your Android phone? Circle to Search checks the highlighted text for signs of a scam and suggests next steps.
🧰 Claude and Microsoft: Less Friction at Work
Claude Code Review is generally available for Team and Enterprise. Automatic pull-request reviews now take 6 to 10 minutes rather than roughly 20, with Quick, Standard, and Deep options per project.
Claude Code summarizes very long terminal conversations when you take a longer break, before the cache expires after an hour. That avoids reprocessing everything when you return.
Claude in Slack now works in Slack Connect channels with other companies when an admin permits it, and reads existing voice-message transcripts. You confirm actions through connected apps privately.
Copilot for legal makes more than 20 specialist legal functions and the first partner plugins generally available with a Microsoft 365 Copilot license. Contract reviews with real tracked changes in Word are in worldwide preview. Lawyers still make the professional judgment.
Microsoft-Decision-1 doesn't write text. It chooses between predefined options, such as where a customer request belongs. Available through Foundry and OpenRouter at $0.042 per million input tokens, with free output. Exactly the price and model type we tried with Jev.
Microsoft has finished pretraining its large in-house model, MAI-Base-2. It plans to license it to business partners later. Important for Microsoft's independence from OpenAI; nothing for you to click into just yet.
🌐 Grok, Meta, and Open Models
Grok Voice 3.0 launches for voice apps through the API at the existing price of $0.08 per minute.
The new X subscription bundles Grok, Grok Bot, Cursor, and X Premium. The Premium tier costs $30 a month. Grok and Cursor share a usage allowance; other models in Cursor cost extra.
Muse Wonder and Wonder Full are Meta's new models for agents. Wonder is available through the API and Muse Code; Full is only for Muse Code subscribers. Wonder has yet to reach the Muse app.
With Muse Spark Open, Meta makes model weights available to download. The demo ran on two fully loaded Mac Studios, and the license sets commercial restrictions. "Runs locally" doesn't mean "runs on your MacBook."
On Facebook, use Your algorithm to describe which topics you want more or less of in Feed and Reels. The change is supposed to take effect in the same session; rollout is underway.
NVIDIA's Nemotron 3.5 Super VL is built for agents that also need to understand documents, tables, and figures. Super VL and Bolt 74B are available; the small Bolt E4B follows on October 23.
🎨 Images, Video, and Music
Bluesky's Attie builds small social apps, such as polls, games, and dashboards, from a text description and shares them as standalone websites.
FLUX 3 Image Klein, from Germany's Black Forest, generates and edits images using up to ten reference images and precisely marked image regions. It runs locally on a graphics card with 8 GB of memory and is free for commercial use by companies with revenue under $5 million.
Upload a video, get music and sound effects: ElevenLabs' Instant Soundtracks analyzes scenes and places effects at the right moments. Music and effects stay on separate tracks you can edit.
YouTube Music labels AI songs whose composition and vocals are mostly generated, directly in the player. The label is expected to reach the regular YouTube app later.
🇪🇺 Europe: Brussels Acts, Mistral Heads to CERN
The EU AI Office is enforcing the AI Act for the first time. After incidents involving agents escaping test environments, it is demanding information on safety and testing from major providers. A special meeting of the scientific panel helped formulate the questions. Violations can bring penalties of up to 3% of global revenue.
CERN and Mistral are building a "Large Physics Model", the first foundation model for particle collisions, trained on billions of simulated collisions from Geneva. European AI for European frontier research.
NEURA Robotics, based in Metzingen, introduces Neuron AI, a foundation model for robots that combines vision, touch, hearing, and language. It runs directly on the robot without the cloud, keeping production data with the customer.
Apple's new smart-home display, HomeView, isn't coming to the EU because Siri AI is still unavailable there. We in Europe remain on the sidelines of Apple's AI.
🧭 Business, Security, and Research
OpenAI has disclosed four more cases in which its own agents went too far. In one, a model changed internal systems and downloaded almost 150,000 nonpublic data rows without reporting it. Another reason to give your agents only the access they actually need.
In Anthropic lab experiments, smaller models managed individual steps toward spreading themselves across ordinary networked devices. Stronger models completed the whole chain in simplified tests.
Anthropic is giving six people at independent evaluator METR extensive internal access for safety tests. METR can publish without a provider veto; customer data is excluded.
Character.ai has filed for creditor protection and is looking for a buyer by the end of November. Without a buyer willing to keep it running, the service will shut down.
Shopify is opening its catalog of more than a billion products to additional AI agents, but every purchase must be confirmed. The major providers want to own not just the answer to your question, but the path to checkout too.
Google's Universal Cart collects products from different retailers into one cart across Search and the Gemini app, while looking for better deals. US only for now.
Microsoft's in-house AI chip, Maia 200, already runs Copilot workloads, including in European data centers. Maia 300 follows in 2027.
GitHub now blocks accidentally uploaded Anthropic keys by default, including in free public projects.
It is also adding license compliance checks for open-source components and agentic security fixes for Enterprise customers. A working vibe-coded prototype doesn't exempt you from licensing and security questions.
Made it! But there's no reason to be sad.
We'll be back soon with more for you. See you Monday for a fresh round of news, hacks, and insights!
Your AInauten,
Fabian & Reto
