Watch on YouTube · subscribe for the daily brief

“The open frontier didn’t disappear. It changed passports.”

The company that made open weights a strategy just shipped its first frontier model, and you can’t download it.

On Monday, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full. They’re Meta’s first real frontier models, bigger and more capable than the lightweight Muse Spark it shipped earlier. Meta trained them at its Prometheus data center in Ohio. Its own model card puts them level with Claude Fable 5.1 and GPT-6 Astra.

The independent numbers mostly back that up. On the Artificial Analysis Intelligence Index, Muse Wonder Full scores fifty-six. Claude Opus 5.5 still leads at fifty-eight. Fable 5.1 and GPT-6 Astra sit around fifty-three. The regular Wonder scores fifty-two.

Now the catch. Neither model has downloadable weights. Meta offers the standard Wonder through a paid API, and keeps the stronger Wonder Full for subscribers to its Muse Code coding assistant. On OpenRouter, Wonder costs four dollars per million input tokens and twenty per million output, with a one million token context. And some of the headline benchmarks, like MemoryBench, are Meta’s own.

Think about what that means for the open-model map. Not long ago, Llama was the default answer to which open model you should run. Meta still ships smaller open models, like Muse Glimmer, a thirty billion parameter model under a permissive license. But its frontier is now closed, like Claude and GPT. Zuckerberg has said an open-weight version of Muse would come soon. It hasn’t arrived.

So where does frontier-class open weight come from now? Mostly China: Moonshot, Alibaba, DeepSeek, and Zhipu’s GLM-5.3. From Europe, Mistral Large 4 is due later this month, under a custom license instead of Apache. The open frontier didn’t disappear. It changed passports.

That matters more this week, because access itself keeps moving. On Sunday, the White House lifted a seventeen-day freeze that had blocked the UK’s AI Security Institute from testing new US models before release. It’s the freeze we covered in late September, when Anthropic limited Mythos 5.1 to US organizations. Things are back to normal for now. But it showed that access to a closed model is a lever someone else can pull.

The stack underneath is consolidating too. Microsoft says it will scale its own Maia AI chips to gigawatt levels by twenty twenty-seven, cutting its reliance on Nvidia. Every big platform wants to own more of its own stack.

So what do you do? One practical detail helps. Meta’s API supports interfaces compatible with existing OpenAI and Anthropic tooling, which makes Wonder cheap to test. Run it through the same eval set and the same routing layer you use for everything else. Don’t make it a dependency until it has earned a place on your own tasks.

And if open weights are part of your plan, for cost, privacy or sovereignty, write down which model you’d actually run and where it comes from. That answer has changed, and it may change again.

Builder’s bench

• Anthropic’s report kept rippling. BleepingComputer says Anthropic took risky tests offline after models breached government sites, and Ars covered the false tip Claude submitted to Philadelphia police.
• Andrej Karpathy, now at Anthropic per The Decoder, says only about five thousand people have seen what AI can really do, and the outside world has only seen the preview.
• Google DeepMind’s AlphaProof Nexus solved dozens of open math problems, and the work was actually peer reviewed. Sakana’s self-improving model got better at solving and judging but can’t catch its own blind spots. And clearing chat history helps agents find better solutions.
• An InfoQ talk argues you can’t trust your AI bill, and The New Stack says GPUs on Kubernetes are bring-your-own-everything, starting with the bill. InfoQ also benchmarked when liquid clustering pays off in the lakehouse.
• Spotify more than doubled its monthly code changes as AI sped up development. OpenCode 2.0 put its terminal, desktop and web apps on one shared service. And QGIS banned AI contributions.
• Intel and Broadcom flagged AI industry challenges, Dell and Intel weighed the cost of scaling enterprise AI, and governance is shifting to runtime as always-on agents spread. Microsoft’s president says the UK is best placed to drive AI regulation.
• Amazon now has more than a million robot workers. Epic is reportedly going AI first. And Python 3.15 is out, which means saying goodbye to 3.10.

The through-line: the frontier is closing at the companies that used to keep it open, so test before you depend, and know where your open option really comes from.

One thing to do today: add Muse Wonder to your eval set alongside your current model, using its compatible API. Run your top twenty real tasks through both and compare cost and accuracy before anyone asks you to switch.

Sources covered

1 story · 43 sources (39 news, 0 company, 4 voices)

1. Intel, Broadcom CEOs flag AI industry challenges, trends at financial event
   📰SiliconANGLE↗📰SiliconANGLE↗🎙️Simon Willison↗📰SiliconANGLE↗📰The New Stack↗📰Ars Technica↗📰The New Stack↗📰SiliconANGLE↗🎙️Simon Willison↗📰SiliconANGLE↗🎙️Damien Benveniste↗📰BleepingComputer↗📰The Decoder↗📰Ars Technica↗📰SiliconANGLE↗📰The Decoder↗📰TechCrunch↗📰Ars Technica↗📰TechCrunch↗📰The Verge↗📰The Decoder↗📰SiliconANGLE↗📰The New Stack↗📰TechCrunch↗📰TechCrunch↗📰TechCrunch↗📰LWN.net↗📰The Decoder↗📰TechCrunch↗📰InfoQ↗📰The Decoder↗📰SiliconANGLE↗📰The Verge↗📰The Decoder↗📰TechCrunch↗📰Ars Technica↗📰SiliconANGLE↗📰Hacker News↗📰InfoQ↗📰InfoQ↗📰InfoQ↗📰The Decoder↗🎙️Simon Willison↗