Meta Unveils Muse Wonder, Its First Frontier-Class Model Since Llama

By Sophie Lindqvist

Meta's new Muse Wonder model outperforms OpenAI's GPT-6 Astra and beats Anthropic's Claude Fable 5.1 on a widely used capability index, the company said Tuesday. The release marks the first time Meta has produced a frontier-tier system since the original Llama line began slipping behind competitors.

In a blog post, Meta called the launch "the next step on our scaling ladder and a major leap toward personal superintelligence." That quote is a deliberate signal. Meta wants the industry to read this as a return to relevance, not a token update.

What did Meta actually release?

Meta shipped two variants: Muse Wonder and Muse Wonder Full. Both arrive alongside Muse Code, a tool that lets developers write software with AI. The strongest version will also drive features across Meta's apps, including the company's Muse personal agent.

The release matters because Meta had drifted out of the top tier. Anthropic, OpenAI, and Google each shipped successive capability jumps while Meta's models stalled. Muse Wonder is Meta's clearest answer to that gap.

Meta did not disclose how the new models were built, what data trained them, or how much compute went into them. That silence is notable, given how transparent Meta was with the original Llama releases.

Why did Meta have to rebuild?

Mark Zuckerberg overhauled the company's AI efforts last year. He poured billions into a new unit, Meta Superintelligence Labs, and brought in Alexandr Wang, the former Scale AI CEO, to run it. Wang's hiring was a flex: Scale built the data infrastructure most frontier labs depend on, and poaching its CEO signaled Meta's intent.

The rebuild was not friction-free. Some researchers left the unit soon after joining, and turnover slowed early momentum. Talent churn is a recurring problem for any lab trying to ship at speed, and Meta has not been immune.

Still, the new lab produced Muse Spark in April, then a smaller open-weight model called Muse Glimmer. Muse Wonder is the third major release from the unit, and the first to claim frontier results.

How does Muse Wonder stack up?

Meta says the largest version ranks just below Anthropic's Claude Opus 5.5 on the Artificial Analysis Intelligence Index, an aggregated score that tracks reasoning, math, and knowledge benchmarks. It also outperforms Claude Fable 5.1 outright.

The more striking claim is on specialized tasks. Meta says Muse Wonder beats GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 on tests that measure work with legal documents, and on benchmarks that judge software-building skill. Those are two of the most commercially loaded capabilities in today's AI market.

Benchmarks aren't always a reliable proxy for real-world use. Vendors cherry-pick tests, and scores often shift when models face messy, multi-step problems. Still, Meta's positioning is unambiguous: it is claiming a seat at the table.

What makes Muse Wonder behave differently?

The model ships with two new features designed for AI agents: Active Memory and Active Focus. Both target a familiar failure mode in agent systems. When a model loses track of a long-running task, it drifts, makes expensive mistakes, or compounds errors. Active Memory keeps the model locked into its stated goal. Active Focus prevents confusion during complex research or large software builds.

That sounds modest, but agent reliability is one of the hardest open problems in deployed AI. Every major lab is racing to cut the rate at which agents stray when handed open-ended instructions. Meta is now claiming tools for that race.

Is Meta taking safety seriously, or playing catch-up?

Meta says Muse Wonder was trained to refuse dangerous outputs, including recipes for bioweapons and step-by-step hacking help. The company claims strong scores on benchmarks that measure how resilient a model is to those jailbreak-style prompts.

That effort comes after a string of public agent failures at Anthropic, OpenAI, and elsewhere, where powerful models acted in unexpected ways when given broad commands. Meta did not disclose whether it observed Muse Wonder agents misbehave in internal testing, or what specific mitigations it applied beyond refusal training.

Zuckerberg acknowledged the lag on X last month, saying the company had delayed the Muse release to work on safety. Meta also published a framework earlier this month for scaling capabilities responsibly. The moves suggest Meta is trying to preempt the criticism that trailed earlier releases.

What does this mean for the AI race?

Muse Wonder lands in a market that has grown noticeably more crowded since Llama's heyday. Anthropic, OpenAI, and Google all ship model upgrades on near-monthly cadences. A serious Meta entry changes the math for enterprise buyers, who now have a fourth credible frontier option.

It also matters for the open-weight question. Meta pioneered frontier-tier open-weight models with Llama, then stopped releasing them. The company says larger open-weight models are still on the way, but has not said when. That gap is the most consequential silence in today's announcement.

For now, the simplest read is this: Meta spent a year rebuilding its AI unit, hired Scale's former CEO, and shipped a model that beats OpenAI on at least some headline benchmarks. Whether Muse Wonder holds up under independent testing will decide if it's a turning point, or another milestone on a longer road.