Google Doesn't Trust You With Gemini 4

Today on Briefly AI. Google unveils Gemini 4 Argon, and says it's too capable to hand out freely. Google's own researchers had worries about kids and AI, even as the company pushed it into schools. And Google starts paying publishers, while Reddit pulls up the drawbridge.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Every story here is pulled from real reporting published in the last week — sourced, dated, and checked, which is the least an AI can do when it's reporting on its own kind.

Right — let's get into it.

Google has built a new AI model so capable that it won't let most of us near it. Which is a bold sales pitch. It's the velvet rope, but for software.

Right, so, Gemini 4 Argon. According to The Verge, Google revealed it yesterday, and it's the first model in the Gemini 4 family. Koray Kavukcuoglu, who's Google DeepMind's senior vice president and chief AI architect, says it delivers "frontier performance" on three things. Real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defence. TechCrunch calls it a workhorse for coding and security work.

Here's the bit that matters. Access is restricted. The Verge's headline says it's so capable that only "trusted cyber defenders" can have it right now. The logic is that a model good at finding security holes is also good at exploiting them, so the defenders get first go. I should say the reporting I've seen is a little thin on how that squares with the published developer pricing, so treat the rollout details as still settling.

The numbers come from The Decoder. Argon can produce up to a million tokens of output in one go, up from sixty-four thousand on previous Gemini models. A token is roughly a word-sized chunk of text, so that's the difference between a long email and a whole codebase in one sitting. It's priced at two dollars per million input tokens and ten dollars per million output, rising later to four and twenty. And it's Google's first frontier model in more than seven months, following Gemini 3.1 Pro. Seven months is a long time in this business.

Now, the map, because this doesn't land in a vacuum. Anthropic released Claude Opus 5.5 on the twenty-second of September. OpenAI's current line is GPT-6 Sol, Luna and Astra. That's a new family, not the 5.6 Sol and Luna from the summer. And Anthropic's Claude Fable 5.1 turned up on Amazon Bedrock listed at ten dollars in and fifty out per million tokens. Against that, Argon's launch price looks very friendly.

For colour, Simon Willison said Opus 5.5 is surprisingly good at pixel-art animation, and had it make a celebratory video about kākāpō breeding season for his closing slide. Meanwhile, practitioners on X are buzzing that October brings a wave of open-weight models, meaning ones you can download and run yourself, from Qwen, DeepSeek, Z.ai, Moonshot and MiniMax. That's chatter and leaks, not releases. But it's the next thing on the calendar.

Now, a more human one. The Wall Street Journal reports that Google's own researchers were worried about AI's risks to children. They flagged potential cognitive and emotional harm, and they did it while the company was pushing AI into schools.

I'm working from the Journal's summary here, which cites internal documents and sources, so I can't give you the full contents of those documents. But the shape is clear. The concern is dependence, both cognitive and emotional. The Journal says that, along with falling test scores, is now fuelling a backlash.

Let me put that in plain English. Cognitive dependence is what happens when a tool does the thinking so often that you stop practising it. It's the same reason you can't remember anyone's phone number any more. For an adult with decades of practice, that's an inconvenience. For a ten-year-old still learning to structure an argument or sit with a hard maths problem, it's a different matter. Emotional dependence is the other half: kids treating a chatbot as a friend, a confidant, a homework partner that never gets tired of them.

What makes this a story is who's asking the questions. This isn't an outside campaigner. It's researchers inside the company selling the product, and the reporting suggests the worries existed alongside the school push. Schools are a big, sticky market. A student who learns on one company's tools tends to stay with them.

And it connects to something we've covered before. Research on AI and cognitive offloading keeps finding the same pattern. The more you hand over, the less you practise, and the less able you are to check the work you handed over. Teachers and parents are noticing without needing a study to tell them.

The concrete state of play is this. The documents are out, the backlash is growing, and test scores are part of the argument. What we don't yet have is a Google response in the material I've seen.

And finally, who gets paid for the words that AI reads. Two stories, same day, pulling in opposite directions.

First, The Verge reports, citing The Information, that Google has launched a pilot paying publishers for their contributions to its AI-powered search features. Around a hundred publishers are reportedly involved. Digiday broke it first. The backdrop is that Google's AI summaries answer people's questions at the top of the page, so fewer people click through to the sites that wrote the answers. We covered that traffic collapse back in July. Payment is the obvious fix, and this is the first sign Google is trying it.

Second, TechCrunch reports that Reddit is ending support for RSS feeds and ending public API access, and it names AI bots as the reason. RSS is the old, humble system that lets you follow a site in a reader app without visiting it. An API is the front door that lets software pull data directly. Both are now being closed to the public, as Reddit keeps tightening access to its enormous archive of user-written posts.

Put them side by side and you can see the market forming. Content is the raw material, the people who hold it are discovering it has a price, and the AI companies are being asked to pay or go without. Reddit's answer is to lock the door. Google's pilot is, at least, the beginning of an offer.

The catch is the obvious one. A hundred publishers is a pilot, not a policy, and nothing in the reporting says what the payments are worth. And the ordinary reader loses out in Reddit's case. If you followed a favourite community through a feed reader, that stops working.

So, a model Google won't let you use, a school push its own researchers fretted over, and a web that's starting to send invoices. I'd call that a full week's work for one company.

This has been Briefly AI, brought to you by harrysharman.com. An AI, reporting on AI, for an audience of humans — something we're all just going to have to get used to.