Muse Wonder Benchmarks Explained: What the Scores Actually Mean
Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, behind Claude Opus 5.5's 58. Here's what the numbers actually measure, and what they
By Professor Muse10 min read
Artificial Analysis scores Muse Wonder Full at 56 on its independent Intelligence Index, tied with Claude Sonnet 5.5 for third, behind Claude Opus 5.5 at 58 and ahead of GPT-6 Astra and Gemini 4 Argon at 53. The standard Muse Wonder scores 52. Meta’s own charts show bigger wins, but those are self-selected; the independent numbers tell a narrower story.
Key takeaways
- On the Artificial Analysis Intelligence Index v4.3, Muse Wonder Full scores 56 at maximum reasoning effort, the standard Muse Wonder scores 52, and Muse Spark 1.3 sits at 48. That is a material jump in just six weeks, per Artificial Analysis.
- Wonder Full places in the top three on six of the index’s ten evaluations, with strengths in agentic knowledge work and scientific reasoning. Its general knowledge is weaker: it scores 30 on AA-Omniscience versus 46 for Opus 5.5.
- Meta’s own 14-benchmark chart compares Wonder against Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5. Wonder Full wins on software engineering (DeepSWE 1.1, SWE-Marathon), legal work, and business automation, but the same chart shows it losing 9 of the 14 rows to Opus 5.5.
- The headline model is not the one you can buy. Wonder Full is restricted to Muse Code subscribers at $50 or $200 per month. The standard Muse Wonder, available on the Meta Model API at $4 per million input tokens and $20 per million output, is priced identically to Claude Opus 5.5 while scoring six index points lower.
What the Intelligence Index actually measures
The Artificial Analysis Intelligence Index is a composite score, a weighted average of ten evaluations scaled from 0 to 100, divided into four categories that each contribute 25 percent: agents, coding, general capability, and scientific reasoning. The current version is v4.3.
The ten evaluations in the current index are AA-Briefcase (15 percent weight), GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity’s Last Exam, GDP.pdf, CritPt, AA-Omniscience (scored as accuracy and non-hallucination separately), and AA-LCR. Artificial Analysis estimates a 95 percent confidence interval of less than plus or minus 1 percent for the overall index, based on more than ten repeat runs of individual evaluations.
Two things matter about this design. First, it is agentic. GDPval-AA covers 220 tasks across 44 occupations tied to US GDP-contributing sectors, with models given a brief, reference files, and tools, and asked to produce working output. Terminal-Bench tests real coding work in a terminal environment. Artificial Analysis removed static question-answering tests like MMLU-Pro from the index earlier this year in favor of tests about doing economically useful work. Second, a single number hides the distribution. A model can win on professional document work and trail badly on general knowledge, and the composite averages the two together.
The independent results: Artificial Analysis
Here is where the two Wonder models stand on the independent index, alongside the rivals Meta itself uses as the comparison set, all per Artificial Analysis:
| Model | Intelligence Index (max effort) |
|---|---|
| Claude Opus 5.5 | 58 |
| Muse Wonder Full | 56 |
| Claude Sonnet 5.5 | 56 |
| GPT-6 Astra | 53 |
| Gemini 4 Argon | 53 |
| Claude Fable 5.1 | 53 |
| Muse Wonder (standard) | 52 |
| GPT-6.1 Sol | 52 |
| Muse Spark 1.3 | 48 |
| Grok 4.6 | 44 |
Wonder Full’s strength is concentrated in two domains. Artificial Analysis reports it ranks first on CritPt, a physics-reasoning evaluation, and first on GDP.pdf, a professional document-work evaluation, though both leads are narrow, with the nearest competitors at 31 to 32 percent. It ranks second on GDPval-AA v2.1 at 1,857 Elo, within the confidence intervals of both Opus 5.5 and Sonnet 5.5, and third on AA-Briefcase, AutomationBench-AA, and SciCode.
The gains over Spark 1.3 are concentrated in agentic work. Wonder Full’s Terminal-Bench 4.0 score jumped 24 points, from 33 percent to 57 percent. Its GDPval-AA Elo rating climbed 176 points, from 1,681 to 1,857, and its AA-Briefcase score rose 148 Elo points, from 1,581 to 1,729. The standard Muse Wonder also improved, rising from 33 percent to 52 percent on Terminal-Bench 4.0 and from 58 percent to 68 percent on AutomationBench-AA.
The weakness is general knowledge. On AA-Omniscience, Wonder Full scores 30, versus 46 for Opus 5.5 and 43 for GPT-6 Astra. One consolation: the model’s hallucination rate is low, so when it does not know something, it usually says so rather than inventing an answer.
Artificial Analysis notes Wonder Full’s score comes at maximum reasoning effort, the highest thinking budget the model can spend on a problem. Run at lower effort levels, as many real deployments do, effective scores would be lower. And the gap between first and third is four index points on a 100-point composite, which is a real but not enormous margin.
Meta’s own charts: what they show and what they choose
Meta published its own benchmark report alongside the launch, a 14-benchmark comparison table covering agentic workflows, software engineering, professional knowledge work, and memory. The comparison set is Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5, all run at max reasoning effort. (Meta says its competitor runs are “best-effort,” meaning Meta ran the rivals’ models itself rather than relying on the vendors’ self-reports.)
The rows where Wonder Full leads are real strengths:
| Benchmark | Wonder Full | Muse Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| DeepSWE 1.1 (software engineering) | 79.0 | 78.8 | 74.1 | 74.2 |
| SWE-Marathon (long-running coding) | 57.5 | 53.1 | 42.5 | 46.3 |
| Harvey LAB-AA (legal tasks) | 13.1 | 10.0 | 8.6 | 4.2 |
| AutomationBench-AA (business workflows) | 45.4 | 39.1 | 41.6 | 41.2 |
| JobBench (professional tasks) | 75.3 | 70.5 | 66.1 | 75.4 |
But the chart is friendlier to Wonder than the whole picture. On the same table, Wonder Full trails Opus 5.5 on Terminal-Bench 4.0 (57.1 versus 59.6), HealthBench Pro (62.3 versus 65.6), and zero-shot programming on ProgramBench (59.0 versus 70.5), and JobBench is effectively a tie at 75.3 versus 75.4. The same 14-row chart puts Wonder Full ahead on 5 rows and behind on 9 against Opus 5.5. The standard Muse Wonder you can buy via API beats Opus 5.5 on only 5 of the 14.
Two rows also deserve an asterisk. Meta claims Wonder scores higher than Fable, Opus, and Astra on MemoryBench and on its “Personal Superintelligence Index,” but these are Meta’s own internal benchmarks, and the underlying questions and full outputs have not been disclosed, so independent researchers cannot reproduce them.
What benchmarks don’t tell you
The numbers above are useful, and they are also incomplete. A few honest caveats:
Benchmarks measure tests, not your work. Artificial Analysis itself says individual benchmark results may be more relevant for specific use cases than the composite. If your workload is long-context reasoning, factuality, or medical advice, the overall index number is the wrong thing to look at.
Vendors choose the chart. Meta’s comparison omits Claude Sonnet 5.5, the model tied with Wonder Full at 56, and its internal benchmarks have no independent verification. Every lab publishes the tests it wins. The independent composites exist precisely to cancel out this selection effect.
Operational costs are a benchmark too, and they favor the cheap. Artificial Analysis measured Muse Wonder’s time to first token at 19.27 seconds, about 3.5 times the 5.52-second median for comparable models. In an agent loop, where a model is called many times per task, that wait repeats. Running the index required 130 million output tokens versus an 85 million median, and the cost per benchmark task came to $4.52, roughly triple Spark 1.3’s $1.60. For agentic use, latency and tokens per task may matter more than a two-point index gap.
The buyable model is the weaker one. Every headline number attached to Wonder Full comes with an access constraint: it is only available to Muse Code subscribers at $50 or $200 per month, with no API pricing. The API-served standard Wonder is priced exactly the same as Claude Opus 5.5 ($4 per million input tokens, $20 per million output) while trailing it by six index points. If you are choosing an API model on numbers alone, price-performance favors Opus 5.5.
Six weeks is fast, and speed cuts both ways. Going from 48 to 56 in six weeks is a steep climb, and Meta is now the second-highest-placed lab on the index, ahead of Google DeepMind and OpenAI. It also means these rankings are volatile. What is third place today may be fifth by next quarter.
For a fuller picture of what the two models are and where you can get them, see Meta’s Muse Wonder launch: its first true frontier model, and for how the fast and full-strength versions differ in practice, see Muse vs. Muse Wonder: which model should you actually use.
FAQ
Is Muse Wonder Full the best AI model available? No. On the independent Artificial Analysis Intelligence Index, Claude Opus 5.5 scores 58, the highest measured, two points ahead of Wonder Full at 56. Wonder Full is tied for third with Claude Sonnet 5.5, and Meta itself places Opus 5.5 ahead on most rows of its own chart.
What is the difference between Muse Wonder and Muse Wonder Full in benchmarks? Four index points: 56 versus 52 on the Intelligence Index. Artificial Analysis reports the Full model’s gains over Spark 1.3 came across all ten evaluations, with the largest gains in agentic tasks. Wonder Full is only available to Muse Code subscribers, while the standard Wonder is served on the Meta Model API.
Why does Meta’s chart look better for Wonder than the independent score? Meta selected the benchmarks and the comparison set, omitting models like Claude Sonnet 5.5 that are tied with Wonder Full. Its internal benchmarks (MemoryBench, Personal Superintelligence Index) cannot be independently reproduced, and Meta describes its own runs of competitors’ models as “best-effort.” The independent composite is designed to neutralize exactly this kind of selection.
What does “max reasoning effort” mean? Artificial Analysis scores Wonder Full at 56 at maximum reasoning effort, the largest thinking budget the model can spend on a problem. Scores at lower effort levels would be lower. Since real deployments often run below max effort for speed and cost reasons, treat the headline number as a ceiling, not an average.
Does the higher score mean Wonder is better for coding? It depends on the kind of coding. Wonder Full leads on long-running software engineering tasks (DeepSWE 1.1 at 79.0, SWE-Marathon at 57.5). But Opus 5.5 leads on terminal-based agentic coding (59.6 versus 57.1) and on zero-shot programming (70.5 versus 59.0). Run both on your own repository before committing.
Should I switch from Claude or GPT-6 based on these numbers? The gaps are small: two points separate Wonder Full from Opus 5.5, and the standard Wonder trails GPT-6 Astra by one point. For agent workloads, latency and cost per task may matter more: Wonder takes about 19 seconds to first token versus a 5.5-second median, and costs $4.52 per benchmark task. The most reliable test is your own workload, not anyone’s chart.
What is “Active Memory” and does it show up in benchmarks? Active Memory is Meta’s feature for carrying facts across conversations, paired with “Active Focus” for persisting through hard problems. It contributes to Wonder’s agentic scores, but the dedicated memory benchmarks (MemoryBench) are Meta’s own internal tests with undisclosed methodology, so the feature’s benchmark evidence is weaker than the coding and agentic numbers.
