Benchmark evidence
Muse Wonder benchmarks: independent scores & evaluation sources
A sourced Muse Wonder benchmark summary separating Artificial Analysis results from Meta’s published scorecard and evaluation methodology.
Last verified · Independent editorial guide
Artificial Analysis lists Wonder (max) at 52 and Wonder Full (max) at 56 on its Intelligence Index snapshot. The table below compares exact model variants. Meta’s own scorecard is a separate source, not a second independent test by this site.
Independent evaluation snapshot
| Model / reasoning | AA Intelligence Index | Input / 1M | Output / 1M | Public API | Source |
|---|---|---|---|---|---|
| Muse Wondermax | 52 | $4 | $20 | Yes | AA |
| Muse Wonder Fullmax | 56 | — | — | No | AA |
| Claude Opus 5.5max | 58 | $4 | $20 | Yes | AA |
| GPT-6 Astramax | 53 | $10 | $50 | Yes | AA |
| Gemini 4 Argonhigh | 53 | $2 | $10 | Yes | AA |
Third-party evaluation · Artificial Analysis· Last verified
Scores: Artificial Analysis Intelligence Index v4.3.2 (higher is better). Prices: evaluator-listed USD per million tokens; rates and tiers can vary. Different reasoning settings are shown explicitly. No tests were run by this website.
An index is not a task-specific guarantee
Editorial synthesis · Cited sources· Last verified
A composite combines multiple evaluations into one number. It can help you shortlist models, but does not tell you whether a particular repository change, document workflow or computer-use task will succeed.
Read the evaluator’s methodology and the exact model settings. This snapshot uses Artificial Analysis Intelligence Index v4.3.2. Wonder and Full use max; the comparison includes the other models’ settings instead of presenting every score as an identical configuration.
Keep published and independently measured results apart
Editorial synthesis · Cited sources· Last verified
Meta’s model card explains which rows use internal runs, third-party reports or provider-reported figures. Some coding rows retain different agent products or harnesses. A number reproduced in an official table does not automatically become an internally controlled result.
We link the model card separately from the independent index. We have not run these evaluations ourselves, and do not average unrelated benchmarks into a new ranking or present estimated scores as measured results.
Build a small evaluation around your own workflow
Editorial synthesis · Cited sources· Last verified
Keep prompts, tool permissions, context, output budget, retries and test cases fixed. Record successful completions, the cost of failures and the human correction time, not just token price.
For coding, verify the produced change with the same tests. For long-running agents, check whether the final requested state was reached. A model that is cheaper per token can still be more expensive per successful task.
Frequently asked questions
Editorial synthesis · Cited sources· Last verified
Why do benchmark numbers differ across websites?
Model versions, reasoning settings, benchmark revisions, tools, harnesses and publication dates may differ. Check the original evaluation rather than comparing isolated numbers.
Did this website run these benchmarks?
No. The independent table links to Artificial Analysis model pages. Meta’s scorecard and methods are linked separately.
Does Full always perform better than Wonder?
A higher aggregate index does not establish that Full wins every task or benchmark. Evaluate your workload and check the individual benchmark rows.
Sources and verification
Provider pricing, model capabilities and access are distinguished from independent evaluations. The date records an editorial check; data is not refreshed automatically.