Meta's Muse Wonder scores 52, Wonder Full 56 on AA Index

Meta Superintelligence Labs was founded roughly 14 months before this launch, and Muse Spark was its first model family. Mark Zuckerberg described Wonder as "a much larger and more advanced class of models that brings us to the frontier." Alexander Wang said Meta continued scaling both pre-training and reinforcement learning since Spark shipped in April.

The launch split availability deliberately: Wonder is the faster frontier model available via the Meta Model API and Muse Code, while Wonder Full is the larger model restricted to Muse Code subscribers. Wonder will also power the Muse personal-agent product, with subscribers getting it first. On launch day, both models drained Muse Code quotas abnormally fast — a bug fixed the same day with a quota reset for all users.

Third-party evaluator Artificial Analysis put Wonder Full at 56 and Wonder at 52 on its Intelligence Index, against Spark 1.3's 48. The gap between Spark 1.3 and Wonder releases was only 6 weeks. Wonder Full improved on all ten index evals, with the largest gains in agentic categories: Terminal-Bench 4.0 went from 33% to 57%, and GDPval-AA v2.1 rose from 1681 to 1857 Elo. Wonder Full ranked top three on 6 of 10 evals, taking #1 on GDP.pdf and CritPt at 33%.

Meta presented plots showing performance on non-post-trained tasks versus post-trained tasks as evidence against benchmark overfitting, which some researchers singled out approvingly. Internally, Alexander Wang said research "starts with evals," with MSL building a Personal Superintelligence Index rather than chasing external benchmarks. Jason Wei summarized the effort as "scaling prevails," noting that much of Meta pivoted to support the model, including business development teams hiring human experts for data review.

The launch drew mixed reactions. Some early users ranked it alongside Opus 5.5 and praised its distinct prose style, while skeptics estimated Meta remains about 3–6 months behind OpenAI and Anthropic. The same day, Artificial Analysis launched a three.js scene benchmark where Claude and Astra models led — a pointed counter to Meta's emphasis on real-world agentic tasks over 3D rendering demos.