Muse Wonder vs Claude shows why the task matters
Muse Wonder has enough evidence behind it to deserve a serious comparison with Claude. The more useful question is which version you can actually use, and which kind of work you want it to finish. The launch results point in different directions once those details are separated.
ByteForward’s new Muse Wonder comparison looks at that split, with strong agent benchmarks alongside a much less convincing reception for some quick visual coding demos. Here is what the published evidence supports.
Check the model name before the score
Meta’s comparison table separates Muse Wonder from Muse Wonder Full. Full scores 59.1 on Meta’s internal Personal Superintelligence Index, ahead of Claude Opus 5.5 at 55.2. The API version scores 54.3.
Full requires Muse Code’s Power Usage or Maximum Usage subscription. See our access guide for details.
Coding benchmarks give different answers
The same Meta table reports Full at 57.5 on SWE Marathon, a benchmark for extended engineering projects, versus Opus 5.5 at 46.3. On ProgramBench, which tests program reconstruction, Opus leads 70.5 to 59.0. These are published evaluation results, not tests reproduced by ByteForward.
The model card qualifies these results. SWE Marathon compares different agents, with Full at xhigh. ProgramBench uses a common agent for clean room reconstruction, requiring at least 95% of hidden tests to pass.
Sustaining a large engineering project and reconstructing a program demand different capabilities. Choose the comparison that resembles your own work.
Independent testing supports the improvement
Artificial Analysis’s launch evaluation gives Wonder and Full Intelligence Index scores of 52 and 56, respectively, compared with 48 for Muse Spark 1.3 six weeks earlier. That provides an independent basis for calling the release a meaningful improvement.
Its coding agent evaluation is encouraging too. Muse Code with Wonder Full and Claude Code with Sonnet 5.5 both display a score of 68. Claude Code with Opus 5.5 displays 66. Those are results for particular model and agent configurations at maximum reasoning effort. Rounded scores should not be stretched into a claim that two products perform identically.
The result is a useful shortlist, rather than a reason to move an entire workflow immediately. Different tools, permissions, prompts and retry budgets can change what an agent manages to finish.
What the demo argument can establish
The accompanying video discusses a lava lamp example and other user reactions as reasons to question how well the benchmark promise carries over to visual coding. Those examples are a prompt for further testing. We have not independently reproduced the demonstrations, and they cannot establish that Meta trained improperly for benchmarks or that the model fails across everyday work.
Meta’s model card lists visual and spatial reasoning, especially 3D, among its limitations.
A convincing comparison would retain the original prompt, model version, reasoning setting, generated files and any follow up fixes. It would also show multiple runs. Without that record, an impressive screenshot and an embarrassing screenshot are both incomplete evidence.
That does not make a poor first result irrelevant. If your actual job is to turn one short brief into a polished interactive scene, the first usable result is part of the product experience. It should be measured directly rather than inferred from a general agent score.
Compare cost per finished task
Artificial Analysis lists $4.52 per Intelligence Index task for Wonder at maximum reasoning effort. That is a weighted cost from its evaluation workload, not a flat price for a chat, a bug fix or a customer project. It also does not price Wonder Full.
For a practical trial, choose a few tasks you already understand. Include a repair in a familiar repository, a fresh build with clear acceptance criteria, and a longer job that requires remembering earlier decisions. Give each system the same starting material and record the model, settings, elapsed time, bill and manual corrections.
Judge the finished work against those criteria. Did the code run? Did the tests pass? Did the requested behavior survive a follow up change? How much checking did you have to do yourself?
Muse Wonder’s launch earns that trial. The evidence is strongest when the question stays specific. Can the version you can access finish your work reliably enough to justify its cost?
