Meta Muse Wonder Tested: Strong Benchmarks, Mixed Demos and Quota Friction

I did not expect to wake up to a frontier-class model from Meta. The Muse Wonder family has two versions, and the published benchmarks put them among the strongest models. A year ago many people had counted Meta out; this is a striking change, although benchmarks are not the whole story.

My Rubik’s Cube test worked: the model built a cube whose faces stay consistent as it scrambles, and the solver actually solves it rather than simply reversing the moves. The project took around four minutes.

Access was more frustrating. I began on the $15 monthly plan but did not have the full version. After upgrading to $50, I ran into 429 quota errors within minutes. I upgraded again to $200 and soon exhausted that quota too. Meta’s Alexander Wang replied that the issue was known and being fixed.

I did manage a few more tests. A browser-based Fall Guys game had flickering and clipping; it was decent but well behind the best versions I had seen from Opus 5.5 and GPT-6.1 Sol. I started an update to see if it could improve.

A Three.js boat on the ocean had attractive water reflections and controls for sunshine, time of day, wind and wave height. But the shadow looked wrong, the boat had odd white material around it, and wind appeared to change water speed rather than wave size. Overall it was far from the best example I have seen.

A 3D watermelon being shot by a bullet was fun to watch but became blurry and did not look hyperrealistic. I wonder whether web technology alone is sufficient for this particular goal and plan to try tools such as Unreal Engine or Blender as well.

At the end of the boat task, the model produced an unexpected message telling me not to investigate its lineage or development. I could not see why the message was relevant to the task, and I do not know what it means. It was an unusual output, not evidence of a particular explanation.

The published results are impressive, although benchmarks differ in usefulness and cannot establish every real-world capability. Meta offers Muse Wonder through its API; the fuller version, Muse Wonder Full, was available through a Muse Code subscription at the time of my test.

The displayed API price for Muse Wonder is $4 per million input tokens, $20 per million output tokens and $0.20 per million cached input tokens. I compare that with GPT-6.1 Sol’s short-context rates of $2, $10 and $0.10 respectively, and Opus 5.5’s $4 input and $20 output rates.

Meta’s comparison places Muse Wonder Full at 59.1 on its Personal Superintelligence Index and 13.1 on Harvey LAB-AA, a legal-agent benchmark. Wonder scores 10 on the latter. I note that Grok 4.7 has a reported 19.6 on the legal benchmark but is absent from the displayed comparison.

On OSWorld, the displayed partial scores are 78.4 for Wonder Full and 72.6 for Wonder; Opus 5.5 leads the shown group at 81.8. Terminal-Bench scores are 57.1 and 52.0 for the two Muse versions, placing the full model near several competitors on that measure.

DeepSWE 1.1 is a coding benchmark I find useful for reflecting developer experience. The chart gives Wonder Full 79.0 and Wonder 78.8, ahead of the other models shown. The small gap between the two versions is interesting, though a single benchmark is not a complete assessment.

On the Artificial Analysis intelligence index, the displayed ranking puts Muse Wonder Full around 56, tied with Sonnet and behind Opus 5.5 at 58. That is an impressive showing. I still prefer using Codex for some work, so a ranking alone does not decide my choice of tool.

The cost-per-task comparison puts Muse Wonder Max at $4.52, while GPT-6.1 Sol Max is shown at $0.72 and several other models cost more. This measures a particular intelligence-index workload, not a universal price for every task. In the intelligence-versus-cost chart, Wonder is strong but relatively expensive.

I also find the MiMo model compelling in the chart because it sits only a few points behind Sol Max at a fraction of its price. A ranking from Theo places Muse Wonder alongside Opus 5.5 in his top tier. Whatever one thinks of individual rankings, Meta’s progress over the preceding year is notable.