Muse Wonder Benchmarks Impress, but Early Demos Lag Behind

What Meta shipped

Meta has released Muse Wonder and a subscriber-only sibling, Wonder Full, which it calls its first frontier AI models. The company is pitching them for coding, legal work, automation and personal-agent tasks, and is emphasizing memory and agent reliability as it rolls them out through its API and its Muse Code product 4. The launch came with a model card and a blog post full of benchmark comparisons 4.

The timing matters. Roughly 14 months ago, Meta had a rough Llama 4 launch and faced a benchmark controversy in April 2025. Since then it has reorganized around Alexander Wang's Meta Superintelligence Labs 1. Muse Wonder is the first real test of whether that overhaul produced a lab that can compete at the top tier.

The independent numbers

The most useful data point is that Meta's claims were not the only ones available on day one. Artificial Analysis posted results the same day as the launch 2. Its Intelligence Index is a composite of ten evaluations, including Terminal Bench, SciCode and Humanity's Last Exam 1. On that index, Wonder Full scored 56, just behind Claude Opus 5.5 at 58 and tied with Claude Sonnet 5.5 14. Standard Muse Wonder scored 52 3. For comparison, GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon all sat at 53, and Grok 4.6 came in at 44 4.

The breakdown is uneven. Wonder Full ranked first on CritPt, a physics-reasoning test, and on GDP.pdf, which measures professional document work 2. General knowledge is a clear weak spot: it scored 30 on AA-Omniscience, against 46 for Opus 5.5 and 43 for GPT-6 Astra 2. It does have a very low hallucination rate, so when it lacks an answer it tends to say so rather than invent one 2.

A second outside evaluator, the Vals index, focuses on real professional tasks. There, Wonder Full ranked fifth at 64.57% accuracy, ahead of every OpenAI model tested. That result prompted a writeup headlined around Meta outranking OpenAI 13.

Meta's own scorecard

Meta's 14-benchmark table shows its sharpest claimed advantages in long coding runs and legal-agent work 4. On DeepSWE 1.1, Wonder Full scored 79.0, compared with roughly 74 for both Opus 5.5 and GPT-6 Astra 4. On SWE-Marathon, which tests extended engineering sessions, Wonder Full scored 57.5 to Opus 5.5's 46.3 34. The widest gap is on Harvey LAB-AA, a legal-agent benchmark, where Wonder Full scored 13.1 and Opus 5.5 scored 4.2 4.

Those figures need a caveat. Meta's own notes say the SWE-Marathon comparison was not run on equal terms. GPT-6 Astra and Fable 5.1 were tested through Codex and Claude Code, while Wonder Full ran at "xhigh" effort 2. Overall, the coding picture is mixed. The model is strongest on very long projects rather than across the board 2.

Where the story wobbles

The launch looks shakier once people start building with the model. Independent testers have described a gap between the benchmark results and early demos. They say 3D rendering and game-building outputs fell short of what Opus 5.5 and GPT 6.1 Soul have produced 1. One example was a Fall Guys-style game with visible flickering 3. Even the more neutral coverage notes that real-world performance remains to be established 4.

This is where concerns about "benchmaxing" come in. Benchmaxing means tuning a model to score well on public tests without matching gains in everyday use. Nothing published so far proves Meta did this. Still, the pattern invites the question: strong results on narrow, long-horizon agent tests, weaker general knowledge, and unflattering creative demos. Meta's own history makes the skepticism louder, given the Llama 4 benchmark dispute 1.

Our read

The fairest reading is that Muse Wonder is a real frontier-class model with a specialized profile, not an across-the-board leader. The independent indexes are the strongest evidence in Meta's favor. Artificial Analysis and Vals both place Wonder Full in the top tier, and neither is Meta's own scorecard 13. Its standout results in physics reasoning, document work and long agentic coding sessions fit Meta's push toward agents and memory 24.

The dramatic leads in Meta's own table deserve more caution, especially where testing setups differed 2. The weak demos suggest the model's strengths may not carry over to visual or creative generation 13. For developers, the practical approach is to try Wonder Full on long-running, document-heavy or legal-agent workflows, where it appears strongest. Pending broader hands-on testing, rivals still look like the safer default for general-purpose and creative work.