Muse Wonder Scores 79, Trails Claude By Two Points

Meta launched Monday two frontier AI systems called Muse Wonder and Muse Wonder Full, with the latter posting a score of 79 on DeepSWE.

Key Takeaways

Meta launched Muse Wonder and Muse Wonder Full on Monday as frontier AI systems for agent workloads
Muse Wonder Full scored 79 on DeepSWE, which tests multi-step software engineering work without human intervention
The larger model ranked second on the Artificial Analysis Index, two points behind Anthropic’s Claude Opus 5.5
Alexandr Wang joined Meta in 2026 to lead its superintelligence effort after Meta’s roughly $14 billion investment in Scale AI

Alexandr Wang, who leads Meta’s superintelligence lab, introduced the series on X, writing that Muse Wonder Full is “our largest and most capable model yet” and that it “rivals the big bois.” Meta’s official AI account confirmed the release, describing both systems as “frontier models, built to deliver state-of-the-art performance” on agent workloads.

Meanwhile, independent benchmark tracker @testingcatalog reported that the larger model scored second on the Artificial Analysis Index, trailing Anthropic‘s Claude Opus 5.5 by just two points.

Muse Wonder’s Index Score And DeepSWE Result

The 79 DeepSWE result is more useful as a question about what the benchmark measures than as a blanket judgment of the model.

DeepSWE specifically tests whether a model can complete multi-step software engineering work without human intervention, a proxy for how well a system functions as an autonomous coding agent rather than a chat assistant.

Based on the figures circulating Monday, that score puts the larger system ahead of most previously disclosed open and closed models on that particular test.

The Artificial Analysis Index aggregates scores across reasoning, coding and agentic tasks into a single comparative ranking of frontier language models.

Also Read: Alibaba’s Qwen Slashes Image Model To 8 Steps, Skips Quality Benchmarks

Its two-point gap behind Claude Opus 5.5 is close enough that routine benchmark revisions could flip the order within weeks, but an aggregate score necessarily reflects the composition and weighting of its component tasks.

Wang’s Push To Close The Gap With Rivals

Wang joined Meta in 2026 to lead what the company now calls its superintelligence effort, arriving after Meta’s roughly $14 billion investment in his prior company, Scale AI.

His mandate has centered on closing a perceived gap between Meta’s Llama model family and frontier labs including Anthropic, OpenAI and Google DeepMind.

The series marks the first model release made public under Wang’s leadership, built explicitly around agentic use cases rather than general chat performance.

The emphasis on “personal agents and professional agents,” the phrase Wang used in his post, signals Meta wants the models judged on task completion rather than conversational benchmarks where it has historically trailed.

What A Public Index Score Can And Cannot Prove

Third-party indices like Artificial Analysis aggregate scores across standardized tasks, giving outside observers a way to compare labs without relying on marketing copy.

That makes the score a question of what those tasks measure, rather than a conclusion about performance across every deployment.

Frontier labs rarely publish comparisons that favor rivals, so a near-top score from an independent tracker carries more weight than a company’s own release language. Even so, the benchmark suite does not capture every real-world agent task a business might run, including long-horizon planning or tool use across unfamiliar software.

Rival labs are likely to respond with their own disclosures in the coming days.

Watch whether the series’ ranking holds once it faces tests outside the DeepSWE and Artificial Analysis suites.