Meta says it's back at the frontier. So why is the model you can actually build on the weaker one? And Meta's not done showing up today. There's also a deal Beijing made it walk away from, and that company just raised big money. You're on AI Daily Briefing. Plus: who's drafting the data-center rulebook, and Epic goes 'AI first.' Muse Wonder up first. VentureBeat, with Carl Franzen:
Independent benchmarking organization Artificial Analysis has already evaluated both models, awarding Wonder Full a score of 56 on its Intelligence Index — tying Anthropic's Claude Sonnet 5.5 and surpassing OpenAI's GPT-6 Astra, Google's Gemini 4 Argon and Anthropic's Claude Fable 5.1, which each score 53. Only Anthropic's flagship Claude Opus 5.5 scores higher among these models, at 58.
Muse Wonder Full gets a 56 on Artificial Analysis. That ties Sonnet 5.5 and beats GPT-6 Astra and Gemini 4 Argon by three. And I can't call it from code. The one on the API is plain Wonder, at 52. Keep the two claims apart. Meta says state-of-the-art on agentic tasks. The independent index says solid, but still two behind Opus 5.5 at 58. And that 56 lives inside a Muse Code subscription Meta controls end to end. Per Kie AI, API Wonder is four dollars in, twenty out per million. That's Opus 5.5 pricing for a 52. Rough procurement memo. And the subscription tier hit a quota bug on launch day, so the good model wasn't exactly reliable either. Same company shipped RoboJEPA with weights and code days ago, so anyone could poke at it. This time, the model getting benchmarked is the one outsiders can't build on. I'd still run the 52 on my own workload, though. Cached input at twenty cents? For long agent loops that keep re-reading the same context, that might be the number that actually matters. Tech Insider, with Sofia Lindström:
NVIDIA has joined AMD, Google, and Meta in backing a new open specification meant to standardize how the industry builds data centers for the AI boom, according to reports from CoStar, Fierce Network, and other outlets tracking the announcement. The effort, built around version 0.7 of what’s being called the Open Data Center Specification, targets the physical layout, cooling infrastructure, and power architecture that AI compute clusters depend on.
Okay, finally a number I can build against. The companion grid spec, which Nvidia, Google, Meta and Microsoft co-wrote, says a facility has to restore near-full load in under two seconds after a grid hiccup. That's a real target. It tells me whether a cluster rides through a disturbance or eats a failed training run. And look at who's holding the pen. Nvidia and AMD sell the chips, Google and Meta are the biggest buyers, and they're the ones defining layout, cooling and power for everyone else. Same Meta from the Muse piece we just did, by the way. Over there it's keeping the stronger model off the API. Here it's writing the rulebook for the buildings. Here's my itch, though. Bitdeer, CoreWeave, all those megawatt announcements. A shared spec probably makes the builds faster and cheaper, which helps the token-cost math. It doesn't tell me a single one of those halls has a signed customer. Right. It speeds up the pour, but the empty-building risk is still there. And Yahoo Finance has them calling it a "global blueprint" at version zero point seven. Drafts get rewritten. I'd wait for 1.0 before anybody rebars a floor around it. From Mahesh at Depth Grid:
On October 8, 2026 AI agent startup Manus said it raised more than $500 million, its first round since Beijing forced Meta to unwind a roughly $2 billion acquisition. Boyu Capital and IDG Capital led the round. Manus has not disclosed its valuation, though earlier reports pointed to $4 billion.
So Manus, the agent startup Beijing pried out of Meta's hands, raised more than five hundred million dollars on October 8. Boyu Capital and IDG Capital led. And the valuation? Undisclosed. That four billion number floating around is from earlier reporting. Manus hasn't put its name to it. Put that next to the Muse Wonder piece from up top. Meta's out pitching agentic tasks the same week the agent company it paid roughly two billion for is raising on its own. And Meta had already started folding in the team and the tech before the planning agency said no. Which is why I'm not sure what this round is pricing. Half a billion with no valuation attached? Feels like investors buying a geopolitical outcome, a Chinese agent shop that stays Chinese, more than a product they've watched hold up over a long task chain. I'd say the geopolitics is the product here. Beijing's stated worry was talent and tech flowing overseas, so whoever funds Manus now is funding who gets to control that stack. Boyu and IDG answered with real money. From Ars Technica:
Epic Games is shifting to an “AI first” development strategy for its collaboration with the Walt Disney Company, letting AI models write all the code and guide all the artwork for an upcoming shooter. That’s according to a Bloomberg report, which cites unnamed people familiar with the project who commented on the significant pivot in internal development at the Unreal Engine maker.
"No more hand-written code." On a Marvel and Star Wars shooter. Somebody still reviews every one of those pull requests, and on a game that size, that's thousands of them. And check the order Bloomberg lays it out in. Understaffed, slower than planned, then the pivot, then delayed indefinitely. That March 2027 date is gone. Which is the tell, right? If the AI workflow were speeding things up, you'd hold the date. Eh, the date was slipping before the pivot. Sweeney's pitch is aimed past this game, at every team building on Unreal. The engine's what he's selling. And the model named in those internal messages? Claude. Anthropic got a showcase without even having to announce it. Sure, but "MORE and BETTER stuff" with no ship date is a roadmap slide. Show me the first build that hits players. From Hacker News:
I guess if any large gaming company is going to fold first, it would be Epic. Pretty horrible experience across the board with them.
"Fold" is a stretch. If anything, Sweeney's leaning in harder. Whether that's better is a separate argument. Launcher grudges aside, Disney's the one more exposed here. Those are their characters riding on a pipeline with no release date. Got feedback, a story idea, or a correction? Email us at aidailybriefing at lantern podcasts dot com. Your notes help us make AI Daily Briefing more useful.
Links to every story we covered are in the show notes, so dig into whichever ones caught your eye. That's AI Daily Briefing for today. Thanks for listening, and we'll be back tomorrow with another briefing. This is a Lantern Podcast.