Two Frontier Models Drop, One Platform Files for Bankruptcy

Grant: I'm Grant, and this is The Z Media Podcast.

Grant: Okay, so this week got loud fast. We've got Meta and Google both pushing out new frontier models within a day of each other, we've got a chatbot platform filing for bankruptcy, an actual arrest at OpenAI over a bug planted in a conference room, and Facebook quietly handing users the steering wheel on their own feed algorithm. Seven stories total, and I'm going to walk through all of them, what happened, who reported it, and the part that actually matters if you're running a content team or a marketing calendar. Let's get into it.

Grant: First up, Meta released Muse Wonder, which TechCrunch and a handful of other outlets are calling Meta's first true frontier model. It's priced at four dollars per million input tokens and twenty dollars per million output tokens through the Meta Model API, that's the application programming interface developers use to plug the model into their own tools. There's also a version called Wonder Full, but that one's locked behind a Muse Code subscription, it's not available through the API at all, so you can't just call it from your own app. Meta says Wonder scores higher than Fable, Opus, and Astra on something called MemoryBench, which is their own memory benchmark. Key word there, their own. And here's the detail that made me sit up, multiple outlets reported the model was trained at Meta's Prometheus data center in central Ohio. Then literally the next day, Google launched Gemini 4 Argon, and this one's available right now through A I Studio, the Gemini platform, and for subscribers in the Gemini app itself through Google A I Ultra and Pro. Google is reporting that Argon ranks number one on something called the Vals Index, which measures performance across finance, coding, legal, and tax work, so it's aimed squarely at professional, multi-step tasks rather than casual chat. Internally Google says Argon helped its own teams raise agentic code submissions by more than thirty five percent inside a tool called Antigravity. And there's a third model story that's smaller but genuinely interesting if you're thinking about on-device tools. Hugging Face and MarkTechPost reported on LFM2.5-24B-A2B, an open-weight reasoning model built to run locally, meaning it thinks through a problem before it answers rather than just spitting out the first guess. It's twenty four billion total parameters but only two billion active per token, and it scored 48.09 on a benchmark called LiveClaw, compared to 38.19 for a competing model that's roughly fifty percent larger, so it's punching above its size class. It's open-weight, available on Hugging Face right now, and it supports basically every major local runtime from day one. If your team is experimenting with agents that need to run privately on a laptop instead of calling out to an API every time, this is worth a look, though the source is clear that it should be paired with search or retrieval for anything knowledge heavy, it's a reasoning engine, not an encyclopedia, so it's better at working through a problem than at recalling fresh facts on its own.

Grant: Now two stories that are less about new tech and more about what can go wrong. The Verge, Wired, and The Rundown A I all reported that an OpenAI engineer, Abdulrahman Alfozan, age thirty five, was arrested and charged with using a device to intercept oral communications, according to a Justice Department release. Here's the part that's wild, OpenAI says they used their own A I models to trace the device back to a specific employee, who was then fired and referred to the F B I. He appeared in federal court after being arrested, and OpenAI reports there's no sign the device touched model weights or customer data, so the damage, as far as they've reported, was contained. For marketers and content teams, the lesson isn't really about OpenAI specifically, it's that if your team handles confidential launch plans or strategy decks in a physical conference room, it might be worth a quick audit of who has access to that space and what's plugged into the walls, the TVs, and the outlets. Second story in this bucket, Character.AI filed for Chapter 11 bankruptcy. TechCrunch, The Decoder, and Ars Technica all covered this one. The company reported assets and liabilities somewhere between fifty and a hundred million dollars, after a string of lawsuits and what the filing describes as struggles to scale. Disney has agreed to buy some of the assets, reportedly for twenty million dollars plus a non-exclusive perpetual license to the underlying technology, and that sale is expected to close in late November or early December. But as of this week, no actual bid or buyer for the rest of the assets had been secured, and the CEO, Colin Adams, said flatly that the platform will be wound down if nothing comes together. Operations are continuing for now, which means anything built on top of Character.AI is still functioning today, but that could change fast depending on how the sale shakes out.

Grant: Two more, and these are the ones your audience will notice directly. The Verge, TechCrunch, and Ars Technica reported that Google is now letting people launch vibe-coded apps straight from the Play Store. It launched in the US this week, and personal apps stay separate from the public store catalog for now, though Google says a feature for sharing with a select group is coming soon. They're also positioning this as a soft on-ramp toward a full professional developer account, with the same trust and safety review process that regular Play Store apps go through. And last one, Facebook is rolling out a feature called Your Algorithm, which The Verge and TechCrunch both covered. It's the same feature Instagram already has, and now it's coming to Facebook, letting users actively add and manage topics that shape their feed and Reels. People can tap interested or not interested from the three dot menu or right under a post, and the system may follow up with questions to refine things further, so it's a conversation between the user and the ranking system rather than a one time setting. Meta says they're leaning more heavily on large language models in ranking now, so users can basically describe in their own words what they want to see more or less of.

Grant: So here's what I'd actually do with all of this if I were running your content calendar this week. One, if you're testing frontier models for anything that needs long memory across sessions, put Muse Wonder on your list, but don't take the MemoryBench claim at face value, run your own comparison against whatever you're using now with your actual content and your actual conversations. Two, go spend real time in Gemini 4 Argon this week while the introductory pricing is live, especially if any part of your workflow is research heavy or multi-step, the cached input discount alone could change your cost math if you're reusing the same reference material over and over. Three, if you've got anything running through Character.AI, audiences, bots, integrations, don't wait until late November to find out if it still exists, start looking at alternatives now while operations are still live and you have time to migrate calmly. Four, if your team holds strategy or launch conversations in a shared conference room, take ten minutes this week and just look at what's plugged in, the TV, the ports, anything that doesn't belong. And five, if you run a Facebook page as part of your content distribution, keep an eye on how Your Algorithm changes reach over the next few weeks, because user-chosen topics could move faster than the algorithm changes you're used to tracking.

Grant: That's the week. Two frontier models, a bankruptcy, an arrest, and a feed algorithm that's about to get a lot more personal. If any of this made you think your own brand's content workflow is held together with duct tape and good intentions, that's kind of our whole thing at Z Media. We build fully automated A I content and C R M systems that write, design, film, publish, and even answer customers for you, so you're not manually doing the thing a system could be doing while you sleep. If that sounds useful, talk with our team, the links are in the show notes, or just call or WhatsApp us at four one nine, five zero zero, four eight eight two. And if you want the key points from weeks like this one delivered daily instead of waiting for the roundup, subscribe to the A I news wire. I'm Grant, this has been The Z Media Podcast, I'll talk to you next week.