Gemini 4, Meta's Muse Wonder, and AI Labels Hit Marketers This Week

Grant: Google just shipped a new flagship model and the price doubles in about six weeks.

Grant: And YouTube Music is now slapping a label on songs made mostly by AI.

Claire: Plus a talking kiosk that wants to upsell your fries. I'm Claire, veteran marketing tech editor, and this is The Week in AI.

Grant: And I'm Grant. This is The Z Media Podcast, eight stories for the week ending Friday, October sixteenth, twenty twenty six. Let's get into it.

Grant: So Google rolled out Gemini 4 Argon this week, its new top model. The Verge and Ars Technica both confirm it's live now for developers through the API, and for Gemini app subscribers on AI Pro and Ultra. Antigravity, Google's coding tool, is Ultra only.

Claire: Okay, so if I'm just on Plus or the free tier, do I even get Argon?

Grant: Nope. The Decoder reports free and Plus users only get Flash models now. Argon fully replaces the old Pro tier at the top. If you want the good stuff, you're paying for it.

Claire: And the pricing. I heard there's a catch there too.

Grant: Big catch. Right now the API is two dollars per million input tokens and ten dollars per million output, that's introductory. Ars Technica says it doubles to four and twenty after November thirtieth. So if you're testing this for content or agent workflows, now's the cheap window.

Claire: Is Argon actually better than what's out there? Google's going to say yes, obviously.

Grant: Google says it ranks first on something called the Vals Index, which covers finance, coding, legal, tax work. But The Decoder points out it still trails Claude Opus five point five in independent rankings. So take Google's number with a little salt.

Claire: So the takeaway for a marketing team is basically, test it now while it's cheap, but don't assume it's the best model on the market just because Google says so.

Grant: Exactly. Run your real documents and drafts through it, measure your actual token use, and decide before the price jumps.

Claire: Okay, my turn. Meta Superintelligence Labs put out two models on October twelfth, Muse Wonder and Muse Wonder Full. TechCrunch and Wired are both calling this Meta's first real frontier-class release.

Grant: Wait, Meta hasn't had a frontier model until now? That feels late.

Claire: That's the framing, yeah. Wonder is available through Meta's API and its coding tool, Muse Code. Wonder Full is locked to Muse Code subscribers only, no API access yet.

Grant: How's it actually scoring against the competition?

Claire: Artificial Analysis has Wonder Full at fifty six and Wonder at fifty two on its Intelligence Index, up from forty eight for the older Spark model. Biggest jump is in agentic coding. But a lot of the comparison numbers come from Meta's own testing, and Wired flags that benchmarks don't always match real-world use.

Grant: What's this going to cost a team that wants to build on it?

Claire: Four dollars per million input tokens, twenty for output. The Decoder says that works out to about four fifty two per task on that Intelligence Index, cheaper than Anthropic's top models but pricier than GPT six Astra. It's also almost three times what the old Spark model cost.

Grant: Any red flags for content teams specifically?

Claire: Weaker factual knowledge scores than top rivals, though it does hallucinate less, which matters if accuracy is the whole job. And Wired notes Meta hasn't said whether its agents misbehaved during testing. If you're piloting agents, run your own tasks before you switch anything over.

Grant: Smaller story, but it matters if you touch music at all. YouTube Music started labeling tracks made primarily with generative AI this week. The Verge and TechCrunch both covered it.

Claire: Where does the label actually show up?

Grant: On the player page for now. Tap it and you get more detail. YouTube says the main YouTube app gets the same labels later, which is the one that actually matters for most branded video.

Claire: How does YouTube decide what counts as AI-made?

Grant: Partners self-disclose as Fully Gen AI, Partly Gen AI, or No Gen AI. Auto-Tune and pitch correction still count as No Gen AI, so it's really about the composition and vocals being generated. If nobody discloses, The Verge reports YouTube will use its own signals to decide.

Claire: Does getting labeled hurt monetization?

Grant: Not on YouTube, according to TechCrunch. AI involvement alone doesn't make a track ineligible. That's actually different from some competitors. Deezer keeps detected AI tracks out of its recommendations and editorial playlists, and TIDAL lets listeners turn off fully AI tracks entirely.

Claire: So what's the actual to-do for an agency using AI music in ads or social content?

Grant: Ask your music supplier exactly how each track was made and keep that on record. Once this expands to the main app, you don't want surprise labels on branded content you didn't flag yourself.

Claire: Next up, SpaceXSI, the company behind X, just merged three separate products into one subscription. TechCrunch and Social Media Today both have this.

Grant: Three products, meaning what exactly?

Claire: X Premium, the Grok chat and agent tools, and Cursor, the coding assistant X acquired back in August. They're calling the whole thing The X Subscription, five tiers, free up through two hundred dollars a month.

Grant: Walk me through the tiers real quick.

Claire: Free, Plus at eight dollars, Premium at thirty, Super at a hundred, Ultra at two hundred. The free tier has pretty limited chat and agent use, and you need Premium or higher just to get the latest Grok models.

Grant: Is this actually a better deal, or just repackaging the same spend?

Claire: X says every plan gives more usage per dollar than what it replaces, and current Premium Plus subscribers pay the same or less. Worth noting that's X's own claim, not independently verified yet.

Grant: So for a brand that's already running a verified account and separately paying for AI drafting tools, what's the move?

Claire: Check your billing settings on X, this rolls out gradually starting in the U.S., and actually compare the usage caps on each tier against what your team really uses before you consolidate anything.

Grant: This one's more about internal tools than campaigns, but it's handy. Google Play now lets you install personal apps built in Google AI Studio straight to your phone, no public listing needed. The Verge, TechCrunch and Ars Technica all covered it this week.

Claire: So these are the vibe-coded apps people build by just describing what they want?

Grant: Right. You build it in AI Studio, and now it installs through Play using your normal Google account. It doesn't show up in search, nobody else can find it, only you can use it. The Verge notes you don't even need to plug your phone into a computer anymore.

Claire: Who's eligible right now?

Grant: U.S. users eighteen and up to start, with Ars Technica saying a few other markets are included too. Sharing with a small group, like friends and family, is coming later, though Google hasn't said how big that group can be.

Claire: Give me a real example of why a marketing team would care.

Grant: Think a quick internal campaign checklist, a content calendar viewer, an event check-in app. Before this, you'd either publish to the store or do a clunky manual install. Now it just lives in your Personal Library on the phone. And if it actually takes off, Google's path lets you move it to a public listing later with the normal developer review.

Claire: Nvidia put out a new open model this week called Nemotron three point five Super VL, aimed at agents that run constantly in the background. The Decoder and MarkTechPost both reported on it.

Grant: Open meaning what, actually open weights?

Claire: Yes, weights, datasets, and training recipes are all published. It's a hundred twenty billion parameters total but only activates twelve billion per token, and Nvidia says it matches their much larger Nemotron three Ultra model on agent and image tasks at about a fifth the size.

Grant: What's it actually good at, in plain terms?

Claire: Document-heavy work. Reading a chart, connecting it to a table somewhere else in the same file, summarizing both without needing a separate vision model. Think receipts, contracts, big batches of PDFs.

Grant: Who's actually using it so far?

Claire: Nvidia names Ramp, EY, Hebbia, CrowdStrike and Salesforce as early users. And there's a smaller version too, a twenty four gigabyte Bolt variant for local machines, out now, with an even lighter four gigabyte voice-focused version coming later this month.

Grant: Is this realistic for a smaller content team to self-host, or is that still enterprise-only?

Claire: Twenty four gigabytes of memory still limits it to developers with decent hardware, honestly. But if you want control over your own data instead of handing it to a provider, this is worth watching, especially once that lighter Bolt model lands.

Grant: Semrush published a guide this week on running SEO audits straight inside Claude, using what's called an MCP server. HubSpot Blog and Semrush's own blog both have the details.

Claire: MCP, for people who haven't heard that term, what is it doing here?

Grant: It's basically a connector that lets Claude read data from another tool, in this case Semrush's Site Audit, which checks a site for over a hundred forty technical issues. Claude can pull a completed audit and work with it right in the chat.

Claire: Can Claude actually run a new audit, or just look at old ones?

Grant: Read-only. It can list audits, summarize issues by how severe they are, show which pages are affected, compare recent audits. But starting a new audit or changing settings still has to happen in the Semrush dashboard itself.

Claire: So what does Claude actually add once the data's in the chat?

Grant: It can rank issues by impact and effort, explain likely causes, suggest fixes, and even say whether SEO, content, or development should own each one. A long audit report turns into an actual assigned to-do list.

Claire: Anything that'll trip people up, cost-wise?

Grant: Every request burns API units, anywhere from a hundred to ten thousand depending on the type, and pulling a past audit specifically costs ten thousand. Several plans include fifty thousand units that reset at renewal, but Semrush is pretty clear, don't pull old audits unless you need to. Check your plan and your unit balance before you build this into a weekly habit.

Claire: Last one. Square launched its first standalone ordering kiosk this week, and it talks. The Verge and Adweek both covered it.

Grant: Talks how, like full conversation, not just tapping buttons?

Claire: Full conversation. A twenty one and a half inch touchscreen, and customers can ask questions, request changes, swap items out loud. Adweek says it can answer things like which dishes are vegetarian or what's under ten dollars.

Grant: Is the AI generic, or trained on each restaurant specifically?

Claire: Trained on that restaurant's own menu, including its popular items and which dishes pair well together. Square's head of food and beverage told Adweek the goal is ordering that feels natural and shows off the best of the place.

Grant: That's interesting, it kind of turns the menu into marketing copy the AI reads out loud.

Claire: Exactly, that's the real takeaway here. How you name and describe dishes, how you group pairings, that's now training material for what the AI pitches customers at the register. It's not just an operations decision anymore, it touches brand and content.

Grant: What's this going to cost a restaurant owner?

Claire: The Verge reports nineteen ninety nine for the hardware, available now to Square Plus or Premium subscribers, plus ninety nine dollars a month per device in software fees. If you run multiple locations, model that monthly cost against any real lift in average ticket size before you roll it out everywhere.

Grant: Okay, so next week, if you only do one thing.

Claire: I'd say it's this. Go test Gemini 4 Argon and Muse Wonder on one real task each, something you actually do every week, before Google's introductory pricing doubles at the end of November.

Claire: Small habit, big payoff later.

Claire: That's the week. New model pricing windows closing, new labels showing up on content you didn't expect, and menus doing double duty as AI training data. Keep an eye on all of it.

Grant: And if you'd rather have a team watching this stuff and building it into your actual marketing and sales process, that's what we do at Z Media. Talk with our team, the links are in the show notes, or call or WhatsApp us at four one nine, five zero zero, four eight eight two.

Grant: If you want this kind of automation working for your own brand, Z Media builds fully automated AI content and CRM systems that write, design, film, publish and answer for you. Talk with our team: the links are in the show notes, or call or WhatsApp us at four one nine, five zero zero, four eight eight two. And for the key points every day, subscribe to the AI news wire.

Claire: And if you just want the key points every day without the deep dive, subscribe to the AI news wire. We're Grant and Claire, this has been The Z Media Podcast, we'll see you next week.