Muse Wonder Benchmarks

Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, while Muse Wonder scores 52. That places them behind Claude Opus 5.5 at 58, level with Claude Sonnet 5.5 at 56 for Full, and close to GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1 at 53 each. Meta’s model card tells a different story: Wonder Full leads on long-running coding, legal work, business automation, and memory. Both claims can be true because they measure different capabilities. That distinction is the most useful takeaway from these results.

Try Apidog today

Below is the full published picture: Meta’s numbers, independent Artificial Analysis results, the specific areas where Wonder leads, where it trails, and which benchmarks cannot currently be reproduced outside Meta. For a broader explanation of the model itself, see the Muse Wonder explainer.

Meta’s model card table

Meta compared Muse Wonder Full, Muse Wonder, GPT-6 Astra, and Claude Opus 5.5 at maximum reasoning effort. The table below follows VentureBeat’s report on Meta’s model card.

Benchmark	What it measures	Wonder Full	Wonder	GPT-6 Astra	Claude Opus 5.5
Artificial Analysis Intelligence Index	Ten-eval composite	56	52	53	58
Harvey LAB-AA	Legal agent tasks	13.1	10.0	8.6	4.2
AutomationBench-AA	Business process automation	45.4	39.1	41.6	41.2
JobBench	Professional work	75.3	70.5	66.1	75.4
DeepSWE 1.1	Software engineering	79.0	78.8	74.1	74.2
SWE-Marathon	Long-running coding runs	57.5	53.1	42.5	46.3
Terminal-Bench 4.0-AA	Terminal agent tasks	57.1	52.0	59.1	59.6
HealthBench Pro	Health questions	62.3	60.5	65.3	65.6

The pattern is clear. Wonder Full performs best when a task is long-running and agentic: it leads SWE-Marathon by 11 points over Opus 5.5, DeepSWE by five points, and Harvey LAB by roughly a factor of three. It trails on terminal work, health questions, and the aggregate intelligence index. Even in Meta’s own table, Opus 5.5 is the stronger general model; Wonder Full is the more persistent one.

Meta states two important caveats. First, competitor runs were best-effort and may not have used optimal settings for those models. Second, some coding comparisons used different tooling, making those rows less controlled than the rest.

Meta’s internal benchmarks: memory and focus

Three figures in Meta’s launch material come from tests that only Meta can currently run.

Benchmark	Wonder Full	Wonder	GPT-6 Astra	Claude Opus 5.5	Claude Fable 5.1
MemoryBench	85.5	85.6	81.4	81.5	79.0
FocusBench	54.2	44.9	47.7	45.8	36.6

MemoryBench places facts inside conversations of 64K to 100K tokens, adds distractors, and changes some facts partway through. The model must recall the current version of each fact. This is the evidence behind Meta’s Active Memory feature.

FocusBench measures whether a model continues pursuing a goal when it encounters blockers. That corresponds to the Active Focus feature. Note that the smaller Wonder model scores below both GPT-6 Astra and Claude Opus 5.5 on FocusBench; only Wonder Full leads. The third internal figure, the Personal Superintelligence Index, has no published competitor comparison.

These benchmarks plausibly test real behaviors, but they also test exactly the behaviors Meta trained for. Until outside evaluators can reproduce them, treat them as evidence of Meta’s training target rather than independently verified performance.

The independent scores: Artificial Analysis

Artificial Analysis evaluated both models using Intelligence Index v4.3 at maximum effort. The composite and component results are below.

Evaluation	Wonder Full	Wonder	Muse Spark 1.3	Notable comparison
Intelligence Index	56	52	48	Opus 5.5: 58; Sonnet 5.5: 56; Astra / Argon / Fable 5.1: 53; Grok 4.6: 44
Terminal-Bench 4.0	57%	52%	33%	Gemini 4 Argon: 57%
GDPval-AA	1857, #2	n/a	1681	Opus 5.5: 1866; Sonnet 5.5: 1838
AA-Briefcase	1729, #3	n/a	1581
CritPt	33%, #1	n/a	n/a	Nearest rivals: 31–32%
GDP.pdf	33%, #1	n/a	n/a	Nearest rivals: 31–32%
SciCode	63%, #3	n/a	n/a
AA-Omniscience	30	25	25	Opus 5.5: 46; Astra: 43; Argon: 42; Sonnet 5.5: 32
Hallucination rate	23%	27%	33%	Gemini 4 Argon: 15%
Coding Agents Index	68	63	54	Claude Code with Sonnet 5.5: 68; Opus 5.5: 66

Three conclusions stand out.

The jump from Muse Spark is real. Wonder Full gains eight points on the composite index in six weeks, and its Terminal-Bench score nearly doubles. Artificial Analysis described this as a material intelligence improvement over earlier Meta models.

Knowledge is the weak spot. Wonder Full scores 30 on AA-Omniscience versus 46 for Opus 5.5, indicating weaker factual recall. Its relatively low hallucination rate, however, means it is less likely to invent missing facts. For retrieval-heavy products, this points to grounding responses with search results or your own documents rather than relying on model recall.

Several leads are narrow. CritPt and GDP.pdf are only one to two points ahead of competing models. Wonder Full’s GDPval lead over Sonnet 5.5 is within confidence intervals. Its Coding Agents Index tie with Claude Code using Sonnet 5.5 is especially relevant for agent-runtime decisions, but that result belongs to Wonder Full, not the API-available Wonder model.

Where Meta and Artificial Analysis disagree

Meta’s model card lists Wonder Full at 59.0 on an Artificial Analysis Intelligence Index row, while Artificial Analysis publishes 56. The discrepancy may come from different index versions, run dates, or Meta’s evaluation harness. Neither party has publicly explained the gap.

When you see 59 quoted, it is Meta’s figure. When you see 56, it is the independent Artificial Analysis result. AutomationBench shows a similar issue: Meta reports 45.4 for Wonder Full and 39.1 for Wonder, while Artificial Analysis reports 71% and 68% on what appears to be a different version of the test. Always quote the source alongside the number.

Tokens and cost behind the scores

Maximum-effort scores come with a token cost. Artificial Analysis measured about 54,000 output tokens per task for both Wonder models, compared with 62,000 for Gemini 4 Argon, 78,000 for Claude Fable 5.1, and 27,000 for GPT-6 Astra.

Input usage is higher still. Wonder used 7.69 million input tokens across the index, while Wonder Full used 6.44 million—two to seven times more than the OpenAI and Google models. As a result, Wonder’s measured cost per task was $4.52: lower than Opus 5.5 at $5.98, but higher than GPT-6 Astra at $3.26 despite Astra’s 2.5x higher list price.

The Muse Wonder pricing post breaks down what these numbers mean for a realistic workload.

What to test yourself

Benchmarks show where to investigate, not what to ship. If you care about the agentic strengths Meta claims, test your own long-running task: a 30-step refactor, a multi-document review, or a tool loop with deliberate failures injected.

In Apidog, use the same prompts and tool definitions against Muse Wonder, Claude Opus 5.5, and your current model. Assert on the structured output and compare token usage for every run.

A useful evaluation record might look like this:

{
  "model": "muse-wonder-1.0",
  "task": "30-step-refactor",
  "passed": true,
  "tool_calls": 31,
  "recovered_from_errors": 3,
  "usage": {
    "input_tokens": 184320,
    "output_tokens": 51240
  }
}

Run enough iterations to see variance, then compare success rate, recovery behavior, structured-output reliability, and total token usage—not just a single passing response.

The tool-call test harness and the non-deterministic agent testing guide cover the setup in detail. You can also download Apidog to compare the models side by side.

Remember that the public API serves Muse Wonder, not Muse Wonder Full. Many benchmark headlines about Meta’s frontier result may describe the model you cannot currently call.

Frequently asked questions

Is Muse Wonder the best model available?

Not on the independent composite. Claude Opus 5.5 scores 58, Wonder Full scores 56, and Wonder scores 52. Wonder Full does lead several specific agentic benchmarks, including SWE-Marathon, DeepSWE, Harvey LAB, and FocusBench.

Which benchmarks can I trust?

Artificial Analysis results are third-party and reproducible. Meta’s cross-model table is Meta-run and uses best-effort competitor settings. MemoryBench, FocusBench, and the Personal Superintelligence Index are currently Meta-only.

How does Muse Wonder compare with Gemini 4 Argon?

Gemini 4 Argon scores 53 on the index versus 52 for Wonder, at half the price. Argon ties Wonder Full on Terminal-Bench at 57% and has a lower hallucination rate: 15% versus 27% for Wonder. The Gemini 4 Argon benchmarks post includes the full comparison.

Does maximum effort turn Muse Wonder into Muse Wonder Full?

No. Effort changes how long the model reasons. It does not turn muse-wonder-1.0 into Muse Wonder Full.

Next step

Choose the benchmark closest to your actual workload. Use SWE-Marathon for long coding runs or AA-Omniscience for factual recall, then run your own task before deciding whether Muse Wonder deserves a trial. The composite index alone will not answer that question.