AI-HOY, AInauts!

Welcome to a new edition of your favorite newsletter.

Another packed and slightly strange week in AI. And not just because of the hidden listening device at OpenAI. Language note: that report is in German.

Two new top models arrived within two days. After what felt like forever, Google is FINALLY back. And Meta is near the front again, too.

We put them all head-to-head so you don't have to.

We also have a fascinating study on who else gets to see parts of your AI chats.

And we'll finish with a practical setup we get asked about at least five times a week.

Here's what's inside:

Let's get into it!

Most of what shapes AI next year is sitting in a research paper today. Keeping up with them is a full-time job, and the abstracts don't help.

TLDR AI is the newsletter that does the translating. Each morning you get the handful of papers worth knowing about, curated by Anthropic and ex-Google engineers, with a plain-English summary of what they found and why it matters.

You come away smarter about AI research in the time it takes to finish your coffee.

Free, and delivered to 1.1M+ inboxes every morning.

Subscribe for free.

Google had been quiet for what felt like ages.

The last big Pro model came out in February. Gemini 3.5 Pro got a big announcement in May. Then it simply never appeared.

Really, things had been quiet since Gemini 3. Everyone else kept leapfrogging each other with new features and capabilities, while Google barely made a sound.

Now the new model is finally here: Gemini 4 Argon.

At the same time, Meta has released its first true top-tier model, Muse Wonder.

Of course we wanted to test them. Should you switch? And when should you use which model?

Visual note: this welcome screen is from our German-language Gemini setup. It says, "See what Gemini 4 can do."

First, a quick look at the boring dashboards.

On the Artificial Analysis Intelligence Index, Google jumps from 30 points for Gemini 3.1 Pro to 53 for Gemini 4. Claude Opus 5.5 scores 58, while GPT-6 Astra also scores 53.

In Arena, where users vote between answers without knowing which model wrote them, Argon is currently in first place. Just ahead of Muse Wonder and Opus 5.5. The gap is small enough to fall within the margin of error.

Funny detail: GPT-6 Astra only lands in 30th place there.

In Peter Yang's comparison, Opus wins at 3D and games, Astra at web design, Gemini at writing, and Muse Wonder with its short film.

So everyone wins somewhere. Here's our everyday test.

Most AInauts aren't looking to build games or run elaborate 3D renders. We wanted to see what happens with the tasks you and we actually deal with every day.

We chose five tasks, all in German, and ran each one three times per model:

For transparency: we built and evaluated the test with Claude. To stop Opus from giving itself a pat on the back, each model only rated the other three models' answers. Blind.

The bottom line: all four solve almost everything.

None invented the book. None invented a study. All found the typo.

And the one real slip-up? It came from our favorite. In two out of three runs, Opus missed a simple arithmetic error (€3,250 instead of €3,450), then happily carried on calculating with the wrong total. Ouch.

More interesting than the scores was how each model answered:

Writing is where it gets really funny.

In the additional self-assessment, every model thinks it's the best. Sounds familiar. 😉

On tasks with a clear answer, the models were honest. On matters of taste, less so. Even the judges disagreed: Gemini gave the Opus posts scores of 10, 9, and 8. Astra rated the same posts 5, 6, and 5.

For us, Astra currently writes best. But we really, really like Gemini's writing style.

Four top models. All four are excellent. Really.

The differences are in personality, speed, and price. Not necessarily capability.

So here are our two lessons:

Our test's big takeaway: Google is back! Good news for anyone who already lives in Google's world.

You can skip Meta. It's good too, but if you already use one of the other providers, you don't need another one.

More generally, learn where each model shines and bring it in when you need it, perhaps for a second opinion. We'll show you the easiest way below.

How we tested: via the API through OpenRouter, with effort set to "high," three runs per task, deliberately excluding coding. This is a small practical check, not a benchmark. The 60 runs plus evaluation cost $8.72.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.

On Monday, in the Musk documentary segment, we asked: "Who are we actually handing our data to when we use Grok and the rest?"

Today we have part of the answer. It's a little unsettling.

There are two separate layers here. First, what the provider itself knows about you. Second, who else gets to read along.

The Muse app mentioned above has built-in long-term memory. Meta calls it Active Memory.

A security researcher showed just how far it goes. He extracted the Muse app's internal instructions simply by asking for them in the chat.

They describe an hourly process that creates a separate page for every person in your life: family, partners, friends, coworkers. Facts, relationships, unresolved topics.

That means information about you can end up in someone else's Muse, even if you've never used the tool yourself. Someone just has to know you.

And according to Meta's help pages, deleting a chat doesn't make Muse forget what it learned from it.

Muse is the extreme case. But all the providers have memory now. In August, our mirror prompt showed just how unnervingly well these tools already know us.

And if you want to take that knowledge with you, you're still stuck copying and pasting. That's why we're still fans of local memory in text files you control. But that's another topic.

Researchers at the IMDEA Networks Institute in Madrid examined nine major AI chat services (read the paper here): ChatGPT, Claude, Gemini, Grok, Perplexity, Mistral, Copilot, DeepSeek, and Meta AI.

They measured traffic in Spain, using EU accounts. Their findings:

Grok was by far the worst. Chat titles and links went to Google advertising, Meta, TikTok, and X.

Chats were publicly accessible by default. Bait links the researchers hid inside Grok chats were opened 70 times from 14 countries.

And sharing a Grok chat sent a screenshot to TikTok too. Like this one:

From the study (CC BY 4.0): a question about a biopsy result is sent to TikTok as a screenshot.

To be fair, not every third party is there for advertising.

With ChatGPT, Perplexity, and in some cases Claude, the data went to error-monitoring and support tools. That's a different category from TikTok.

But Claude didn't come out looking great either. An email address reached its support tool even when cookies were rejected. And the Android app sent location data to a fraud-detection service.

EU data-protection authorities have already been informed. Business plans were not tested in the study.

Yes, we all know on some level that AI labs can access our chats and store things somewhere. None of this started with social media.

We use memory features every day. They're useful.

But they're still a profile a company keeps on you. You only see part of it. And sometimes it includes people who were never asked.

Data making its way to advertising tools isn't exactly great. But it was probably only a matter of time before chat became part of the same tracking machine as the rest of the internet.

We're not panicking. We are being more deliberate about what we type, and taking another closer look at fully local alternatives.

Let's finish with something you can put into practice right away.

We've shown this before, but questions keep coming in. So we wanted to revisit it.

Remember the question at the top: how do you use other AI models' strengths without paying for four subscriptions?

Longtime readers know we do this through OpenRouter. That's also how we ran our test.

Visual note: the OpenRouter and connector screenshots below show our setup; some interface labels and chat text are in German. The instructions are translated below.

One account, all the models. You only pay for what you use.

It's incredibly useful to bring in a second model for a quick sparring session, whether you're working on copy or a complex analysis.

Here are two ways to get OpenRouter into your main chat.

At openrouter.ai/chat, you can ask one question and see several models' answers side by side.

For us in Germany, Austria, and Switzerland, this is also the easiest way to try Muse Wonder at all. And it lets you use Gemini 4 without a Google subscription.

The drawback of OpenRouter Chat is that supplying context isn't easy.

So it makes more sense to give your main AI agent the ability to call in other OpenRouter models when it needs them.

If you use Claude or ChatGPT, tell it you want to use other models through the OpenRouter MCP connector. Sign-in uses OAuth; you don't need to create an API key yourself. OpenRouter provides its own server for this.

In Claude: Settings → Connectors (under Customize in some versions) → + → "Add custom connector." Enter https://mcp.openrouter.ai/mcp as the URL, leave the OAuth fields blank, connect, and approve access in your browser on OpenRouter.

In ChatGPT: in Plugins, or at chatgpt.com/plugins, click "Add" → "Create custom MCP server." Use the same URL and sign in with OAuth.

Then just write: "Ask Gemini 4 Argon through OpenRouter what it would improve about this proposal." The answer arrives in your chat, with whatever context your main chat passes along.

Nobody needs four subscriptions.

One daily driver and an OpenRouter account with $10 are enough to get a second opinion on important work. And to test each new model yourself on day one.

Then you can make up your own mind instead of letting benchmarks and YouTube thumbnails make you nervous.

And as you can see in the screenshot above, OpenRouter doesn't just offer language models. There are images, video, and more.

You just need to know which models exist and where they're useful. Then bring them in for the task at hand.

Made it! Another long one. At least we didn't need four subscriptions to write it. 😉

Thanks for reading!

Reto & Fabian from AInauten