Muse Wonder and Muse Wonder Full: Meta's First Frontier Model

Six months into this series, Meta finally shipped the thing the first thirty articles kept gesturing at without quite having yet: a model built to compete at the frontier, not just to run an agent fast. On October 12, 2026, Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full, which the company's own announcement calls its first "frontier model." That's a specific, deliberate word choice — Muse Spark, the model that ran the original September launch, was built for low latency and real-world agentic work, not for topping leaderboards. Wonder is Meta's attempt to do both at once.

What Meta claims, and what's actually been checked

Meta's model card says Wonder and Wonder Full deliver "comparable benchmark scores" to Fable 5.1 and GPT-6 Astra, and that Wonder beats all three — Fable, Opus, and Astra — specifically on a memory-retention benchmark called MemoryBench. Those are Meta's own numbers, from Meta's own model card, and worth treating as a starting claim rather than a verdict.

The more useful number, because it comes from an independent evaluator rather than the company being evaluated, is Artificial Analysis's Intelligence Index: it puts Muse Wonder Full at 56 and Muse Wonder at 52, with Wonder (Max) landing at #9 out of 214 models tracked. That's a real, externally-produced data point, and it's the one worth citing if you want a number that isn't marketing copy. It doesn't settle the "comparable to Fable and Astra" claim either way — Artificial Analysis's index is a composite across many benchmarks, not a head-to-head on the specific tasks Meta highlighted — but it does confirm Wonder is a genuinely frontier-tier model by at least one outside measure, not just a confident press release.

Reading the benchmark number correctly

It's worth being precise about what the Artificial Analysis number does and doesn't establish. The Intelligence Index is a composite score built from many underlying benchmarks, weighted and combined into a single figure that lets models from different labs be ranked against each other on roughly comparable terms. A score of 56 for Wonder Full and 52 for Wonder, with Wonder (Max) landing at #9 of 214 tracked models, tells you Wonder is competitive with other models currently near the top of that list — a genuinely frontier-tier result by that one measure. It doesn't tell you whether Wonder is actually better or worse than Fable, Opus, or Astra on any specific task a real user cares about, like drafting a contract clause or debugging a production incident. Composite indices are useful for a quick read of where a model sits in the field; they're a poor substitute for task-specific evaluation, which is exactly the kind of detail Meta's own model card and research blog are supposed to provide and which outside evaluators haven't yet independently reproduced for every claim Meta made.

Two models, two different jobs

Wonder and Wonder Full aren't the same product wearing different pricing tags. Wonder is the model available now through the Meta Model API and through Muse Code, Meta's coding product. Wonder Full is restricted to Muse Code subscribers — it isn't served through the API at all, a distinction we unpack further in Why Wonder Full Skips the API. Neither model is built into the consumer Muse app yet; Meta says both will "soon" power more of what muse.ai does, with subscribers getting access first and no firm date given for everyone else.

The capabilities Meta is actually highlighting

Setting benchmarks aside, Meta's own framing for Wonder centers on four things: expert-level performance in law, finance, and health; frontier-level coding for building real software like games and business apps; Active Memory, which carries relevant information across conversations rather than starting fresh each time; and Active Focus, which is meant to keep the model working toward a goal through setbacks instead of giving up at the first error. Active Memory is the capability Meta's own researchers have talked about most directly — Kavya Kopparapu, part of the team behind the release, described the goal as making the model "remember what's important to you over days, weeks, months, and beyond," which lines up with Meta's MemoryBench claim and is worth watching for whether it holds up once Wonder is actually running inside the consumer app rather than just the API. We go deeper on that specific claim in MemoryBench and Active Memory, the pricing implications in 4 Dollars In, 20 Dollars Out, and the Active Focus claim specifically in Active Focus.

How this compares to the launch this series started with

Six weeks ago, when this series opened with What Is Muse?, the entire product ran on Muse Spark — a model Meta explicitly described as built for low latency and real-world agentic work, not for benchmark leadership. That framing made sense for a launch built around speed and distribution: Muse arrived into an installed base of billions of Instagram and Facebook users, and the product's early advantage was reach, not raw model quality. Wonder changes that calculus. It's the first time Meta has paired its distribution advantage with a model it's willing to put up against Fable, Opus, and Astra on their own terms, rather than simply running the fastest agent it could ship into the biggest audience available. Whether that combination — frontier capability plus consumer-scale reach — ends up mattering more than either half alone is the open question this entire continuation batch is trying to work through, one claim at a time.

The context nobody at Meta is saying out loud

This is the first frontier-tier release since Alexandr Wang restructured Meta's AI efforts in August 2025. In the time since, Meta shipped Muse Spark and the Muse agent itself, but it had conspicuously avoided the kind of state-of-the-art model release that OpenAI and Anthropic treat as a routine cadence. Wonder is Meta's answer to that gap — not a bigger Muse Spark, but a statement that Meta intends to compete on raw model capability, not just on agent product design.

Whether that statement holds up depends on things this article can't yet answer: whether Wonder Full's benchmark advantage survives independent red-teaming, whether Active Focus actually recovers from errors the way Meta describes rather than just retrying the same failed approach, and whether "soon" for consumer rollout means weeks or months. Meta has a specific, trackable safety report "coming soon" that hasn't published yet either — we cover exactly what's still unverified there in The Safety Report That's Coming Soon. For now, the honest summary is that Meta shipped a model with at least one outside benchmark backing up the "frontier" label, built on infrastructure — Meta's Prometheus data center in central Ohio — purpose-built for exactly this kind of release, which we look at in Built at Prometheus.

Part 31 of 46 in the Meta Muse series. Previously: Muse Six Months Out: The Questions Meta Hasn't Answered. Next: Why Wonder Full Skips the API: Muse Code as the New Moat. Browse more builder insights or explore AI skills at aiskill.market.