Two of the biggest companies in tech spent Monday showing how much of the AI stack they now want to own. Meta shipped its first frontier models, trained in its own data center. Microsoft unveiled a third-generation chip it wants to run its own inference. Meanwhile, according to Bloomberg, Epic Games told developers on its Disney shooter that the code would no longer be written by hand. They'd be using Anthropic's Claude.
Own the stack, or rent someone else's and bet the studio on it. That split ran through almost everything worth reading today.
Meta prices Muse Wonder like Opus and keeps the Full model for itself#
Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, its first real attempt at a frontier model since the AI reorg under Alexandr Wang. Both take text, image, audio and video in, put text out, and have a context window of roughly one million tokens. Meta says both were trained at its Prometheus data center in Ohio. Neither comes with downloadable weights, which is a notable shift for the company that made Llama the default open model.
The numbers, mostly via VentureBeat's breakdown:
Artificial Analysis Intelligence Index: 56 for Wonder Full, 52 for Wonder. Claude Opus 5.5 sits at 58, Sonnet 5.5 at 56, GPT-6 Astra at 53.
Muse Wonder API (muse-wonder-1.0): $4 input / $20 output per million tokens, $0.20 cached input. That's exactly Opus 5.5's list price.
On Meta's own table, Wonder Full leads on SWE-Marathon, its long-running coding benchmark (57.5 vs 46.3 for Opus 5.5), and trails slightly on Terminal-Bench 4.0 (57.1 vs 59.6).
Wonder Full is not in the API at all. It's only available through Muse Code subscriptions at $50 or $200 a month.
That last point is what I'd focus on. Meta has a model it thinks is competitive at the top, and it chose to put that one behind a consumer-style subscription rather than an API. Contributor mode is on by default for those Muse Code plans, which means Meta may use your prompts, code, files and outputs to train future models. The pay-as-you-go Wonder API runs in standard mode, where Meta says it doesn't train on what you send. So: the best model costs you a flat fee plus your data, and the cleaner deal gets you the weaker model.
Meta's benchmark card also mixes third-party results, vendor numbers and internal evals, and some coding comparisons used different harnesses. I'd treat the SWE-Marathon lead as a reason to run your own long-horizon tasks against it, not as a verdict. The API speaks both OpenAI- and Anthropic-compatible interfaces, so trying it costs about an afternoon.
Maia 300 is Microsoft's hedge against its own GPU bill#
Microsoft announced Maia 300, its third in-house AI chip, and said the Maia family will become the primary inference system for its own products: Copilot, GitHub Copilot, Foundry. Microsoft claims 40% more tokens per megawatt than leading GPU systems, plus 36% more compute and 2.5x the HBM bandwidth of Maia 200, per Techzine. Production deployments start in 2027, with multi-gigawatt scale as the stated goal.
Every number here comes from Microsoft, and nothing says Azure customers can pick Maia for their own workloads. For builders, that's fine. You won't touch this chip directly. You'll touch it as a GitHub Copilot price that Microsoft can hold flat or cut, because the token economics underneath are no longer set by Nvidia's margins. Meta did the same thing one layer up with Prometheus. Owning the expensive part is how you get to set the price of the cheap part.
Epic rents its whole pipeline instead#
Then there's the other camp. Ars Technica, summarising Bloomberg's report, says Epic delayed its unannounced Disney shooter with Marvel and Star Wars characters, originally slated for March 2027, and moved it to an AI-first model. A director reportedly told staff there would be no more hand-written code, art would come from full AI workflows, and developers should use Claude with humans steering. Epic's comms director Jake Jones told Engadget the company is "embracing AI-accelerated development" with its developers still leading creative work. Staff reportedly found the directive shocking.
I'm torn on this one. I use coding agents every day, and "no hand-written code" is a slogan, not an engineering plan. Last week's coverage of agents producing more code without producing more software applies directly here. A delayed AAA project with a fixed licensing partner is about the worst place to discover where that ceiling is. And unlike Meta or Microsoft, Epic owns none of the stack it's now dependent on.
The rules aren't keeping pace with either camp#
Microsoft's other headline came from its president. On a BBC podcast, Brad Smith called for urgent AI regulation and argued the U.K., not Washington or Brussels, is best placed to lead it. He floated mandatory breach notification when rogue AI agents compromise a system, outside evaluators signing off before release, and kill switches. On voluntary pledges, he didn't hedge:
"It's the companies that are not as responsible that you need to worry about." — Brad Smith, via TechCrunch
It's an easier argument to make the same day your company announces it'll own more of its inference stack. Still, the breach-notification idea is the one I'd actually want as a builder. It's narrow, it's concrete, and it treats agent incidents like security incidents, which is what they are.
The smaller story shows where that oversight is missing. Ambi, founded by former Plaud exec Peter Mo, raised $26 million for iPhone, Mac and Watch apps with an optional always-listening mode. The company says it doesn't store recordings and encrypts data, without saying how, and tells users it's on them to inform the people around them. That's the agent-privacy problem in a pocket-sized form: the people being recorded never agreed to anything.
My bet: within a year, the companies that own their silicon and their training clusters will be the only ones whose top models you can't simply call by API. Meta just went first. If Wonder Full's long-horizon coding numbers hold up in independent testing, the question for teams like Epic's won't be which model to rent. It'll be whether the best model is for rent at all.