Good morning and welcome. It is Tuesday, October 13th, 2026. You are listening to Not a Llama, the daily podcast where we cut through the AI noise and tell you what actually matters. Today we have three big stories. First, Meta launches Muse Wonder, its first real shot at challenging Anthropic and OpenAI. Second, we look at the numbers that decide whether that challenge is real. And third, Matthew Berman breaks down the whole launch. Let us dive in.

Meta just threw down a glove. After years of chasing the big two, the company has landed what it calls its first genuine frontier model family — and it is aiming straight for the top of the pack.

Here is the picture. On Monday, October the twelfth, Meta's Superintelligence Lab released two new models, Muse Wonder and Muse Wonder Full. According to TechCrunch and The Verge, this marks Meta's first true frontier release, built on the earlier lightweight Muse Spark model and the Muse agent.

Wonder is available through the Meta Model API. Wonder Full is more exclusive, restricted to Muse Code subscribers and not served through the API. Subscription access reportedly starts at twenty dollars for personal users and fifty for developers, with the exact pricing printed in today's notes.

On the numbers that actually matter, Meta says its internal testing claims Wonder matches Fable 5.1 and GPT-6 Astra on complex agentic and coding tasks. The company also claims a win on memory, pointing to what it calls active memory for retaining information over long stretches. Wonder pricing sits around four dollars per million input tokens and twenty dollars per million output tokens, which TechCrunch describes as roughly equivalent to GPT-6 Sol or Opus 5.5.

The models were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year. Meta has not released parameter counts or context length.

There is healthy caution in the room. The Verge notes it is unclear how these models compare to Google's Gemini 4 Argon. Ground News says external testing remains necessary, and Nathan Lambert on Bluesky called the release back at the frontier — while stressing that independent evaluations are still needed. Investors reacted quickly, with Meta's shares rising about four percent in afternoon trading.

The question now is whether the hype holds up — once the experts start testing for real.

We covered Muse Wonder a few days ago, but something new has surfaced — and it is worth your attention.

Here is the situation. Meta has published its own benchmark figures for the new Muse Wonder reasoning model. On one unnamed test, Meta lists eighty five point six percent, compared to eighty one point five percent for Claude Opus 5.5 and eighty one point four percent for GPT-6 Astra. On the SWE-Marathon coding challenge, Meta reports fifty seven point five percent against forty six point three percent and forty two point five percent. Meta also says it leads on DeepSWE — though it gives no number.

Now, according to posts on Reddit, the conversation has shifted. People are asking whether they should actually believe these claims. The scores come from Meta itself, not from any independent lab. One poster even questioned whether to trust Zuckerberg directly. Another points to a single user screenshot that nobody can verify.

So the real question is not whether the numbers look impressive. It is whether anyone outside Meta has confirmed them. Until then, treat these figures as claims — not proof. The pricing and full details are waiting in the episode notes.

Here is an update on the Muse Wonder launch.

Matthew Berman has published a hands-on walkthrough of the models, and it runs about eleven minutes. The value is context. He walks through what the Wonder models can actually do — versus what the marketing promises. That is the honest gap worth paying attention to.

Two models made it out of the family. Muse Wonder is reachable through the API. The other one — Muse Wonderful Full — is locked behind a Muse Code subscription. Berman calls them massive frontier models.

He also flags some rough edges. He hit quota errors, those four two nine responses, but it looks like a single-user experience. Alexander Wang, who heads AI at Meta, confirmed the issue was known and being fixed. There was also a mysterious leave your lineage pure message — that no one has explained.

On the numbers side, he compares the models against Opus 5.5, GPT 6.1 Soul, GPT 6 Astra, Fable 5.1 and Gro 4.7. Those names are not independently verified here, so take them with a grain of salt.

The full pricing and benchmark details are waiting for you in the episode notes.

Here is a fresh wave of open-weights models landing today — and this one is a real surge. Let us get into it.

First, Decision 3.0 from the vLLM Semantic Router team. It is a decision model, meaning it returns probabilities for choices, yes or no answers, and scores, with no text generation. It reads text, images, and video, and the 27B version tops both Jev Decision Index boards according to the developers. Their numbers are self-reported, so keep that in mind. It comes in sizes from 0.6B to 27B, and it is Apache licensed.

Next, MiniCPM-V 4.7 from openbmb brings the first mixture of experts vision model in the series. They claim it runs smoothly on-device at more than thirteen tokens per second on the latest iPhone, and that it beats Qwen3.6 in less inference time. Again — that is the developer speaking.

And finally, GLM-Legal opened a new legal vertical, and it is already drawing serious attention online, with hundreds of points and comments lighting up Hacker News. The details are thin, but the interest is clear.

Here is an update on that OpenRouter story. A fresh round of repricing just landed on the platform, touching DeepSeek, Z.ai, Moonshot, Google, MiniMax, OpenAI, and Meta. The one headline worth holding onto is DeepSeek's Pro model, whose output price has dropped by more than eighty-five percent — dramatic. That is a dramatic cut, and it is the kind of move that reshapes how people use that model for heavy lifting.

Elsewhere the picture is mixed. Kimi's Latest sees its output price fall from sixteen and a half down to ten dollars. But several smaller models have gone up. MiniMax M2.7, OpenAI's gpt-oss, and Meta's Llama 4 Scout all rose, sometimes by a wide margin. Input costs moved too, with a few DeepSeek entries nudged upward.

We are not reading out the numbers here. The full breakdown of every input and output price is sitting in the episode notes, so go there if you want the exact figures. The short version is that output is getting cheaper for the big models while a batch of flash models has quietly crept higher. That is worth keeping an eye on as these platforms keep adjusting.

Microsoft just announced its third-generation in-house AI chip, the Maia 300 — and the message is clear. This is the company's bet on reducing its dependence on Nvidia hardware. According to The Verge, the chip is built to be the primary AI inference system for Microsoft's products and services. It will start rolling out to production workloads next year, with Azure Maia deployments scaling up to multi-gigawatt sizes starting in 2027.

Microsoft's own numbers are the headline here. Executive VP Scott Guthrie says the Maia 300 is expected to deliver forty percent more tokens per megawatt than leading GPU-based systems. In a full ninety-six-accelerator configuration, Microsoft claims up to seven times the rack-scale token throughput of its predecessor, the Maia 200.

The chip is built on TSMC's newest process node and uses a new chiplet architecture that separates compute, memory, and input or output functions. Microsoft says this design increases the silicon area dedicated to performance. The Maia 300 succeeds the Maia 200, which launched in January and is already powering Copilot, GitHub Copilot, and Foundry.

All of these figures come straight from Microsoft and have not been independently verified — the goal of catching up with Google DeepMind, OpenAI, and Anthropic remains Microsoft's stated ambition, not a measured result.

Today the big takeaway is that Meta's Muse Wonder has arrived — looking like its first genuine frontier challenge to Anthropic and OpenAI. We covered what the numbers mean, Matthew Berman's breakdown, the open-weights surge, fresh OpenRouter price changes, and Microsoft's Maia 300 chip to cut Nvidia dependence. Thank you for listening, and a new episode arrives tomorrow morning.