Meta Muse Wonder Benchmarks Explained: All 14 Tests Deconstructed

👁 117 reads

Meta showcased Muse Wonder with a 14-test benchmark comparison where most people will barely recognize two names. Below is a straightforward breakdown of every single one: what it actually tests, Muse Wonder’s scores (according to Meta), and what caveats to keep in mind before assuming the model “wins.” In short: 3 out of 14 are Meta’s internal evaluations, several rows lacked equal testing conditions, and in greenfield programming and computer control, Meta still trails Claude Opus 5.5. For a complete model breakdown, see the Muse Wonder guide.

All 14 benchmarks in one table

Source: Meta model card and benchmark tables dated October 12, 2026. “Evaluator” based on methodologies detailed in the model card.

Personal workflows: PSI and MemoryBench

Personal Superintelligence Index is the headline metric of the launch and simultaneously the least independent. Meta compiled it from seven benchmarks: Agents’ Last Exam, AutomationBench-AA, FocusBench, JobBench, GDP.pdf, VBT, and DRACO. Two of these (FocusBench and VBT) are internal proprietary suites. Selecting the ingredients was entirely Meta’s decision, so the index simply shows that Wonder Full leads on a specific collection of benchmarks Meta deemed essential. Think of it like a decathlon: coming out ahead overall does not mean winning every individual event.

MemoryBench evaluates whether the model remembers crucial facts across long conversational sessions (64,000-100,000 tokens), even when contradictory or updated information was presented along the way. That is vital for an autonomous agent tasked with retaining user preferences, but the test is also proprietary and graded by an LLM-as-a-judge. Interestingly, base Wonder scored 0.1 points higher than Wonder Full here. I evaluated a similar mechanism myself across 80 board meeting minutes where earlier decisions were later overturned - all five models passed, as detailed in my 23-task practical evaluation.

Enterprise workloads: legal, finance, healthcare, and office work

• Harvey LAB-AA - Meta’s largest recorded margin: 13.1 points versus Claude Opus 5.5’s 4.2. Raw scores are low across the board because a submission only passes if the generated brief satisfies all criteria without a single hallucination. Version 1.1 added strict hallucination penalties, making comparisons to v1.0 invalid.
• Vals Finance Agent V2 - Analyzes company financial filings and market datasets. Wonder Full edges out Fable 5.1 by just 0.2 points, effectively a tie.
• HealthBench Pro - Answering 525 professional clinical questions without external tools. Meta loses here: 62.3 versus 65.6 for Opus. Anthropic’s scores were pulled directly from Claude system cards, not benchmarked independently by Meta.
• GDP.pdf - Document comprehension on provided PDF files, evaluated by Surge. GPT-6 Astra takes first place at 34.2.
• JobBench - 65 practical tasks spanning 35 office professions. A dead heat between Wonder Full and Opus (75.3 to 75.4).
• AutomationBench-AA - Workflow automation across simulated enterprise software, graded strictly on whether the application’s final state matches the target state. Wonder Full scored 45.4 versus 41.6 for its closest competitor.

Desktop control: OSWorld V2.1

OSWorld puts models through 108 practical tasks on an actual Ubuntu desktop: the model observes screenshots, moves the cursor, clicks, types, and scrolls. Meta presents two distinct figures, and you need to look at both. Under partial credit (rewarding individual completed milestones), Wonder Full achieves 78.4. Under binary evaluation (full completion or nothing), it drops to 39.4. In practice, the model completes roughly one out of three multi-step complex tasks from start to finish. In this category, Meta trails Opus 5.5 (81.8 partial) and Fable 5.1 (80.7 partial, 43.1 full completion).

Coding benchmarks: four tests, four different challenges

Meta takes the lead when maintaining existing codebases on marathon tasks, but falls behind when building programs from scratch or running terminal workflows. Two important caveats from the model card: on SWE-Marathon, results for GPT-6 Astra and Fable 5.1 were run using their proprietary agent frameworks (Codex and Claude Code), while Wonder Full used maximum reasoning (“xhigh”). Meta itself notes this was not an apples-to-apples environment. In ProgramBench, a passing grade requires clearing at least 95% of hidden unit tests.

The only independent composite: Artificial Analysis Index

Artificial Analysis runs 10 independent benchmarks directly (including GDPval-AA, Terminal-Bench 4.0, Humanity’s Last Exam, and SciCode) and normalizes them into their 4.3 composite index. Checking their leaderboard on October 12 showed Opus 5.5 at 58, Wonder Full at 56 (matching Claude Sonnet 5.5, which Meta omitted from its own visual charts), and standard Wonder at 52 - tying GPT-6.1 Sol. This composite is the single metric I find most trustworthy. It represents normalized index points, not raw percentage accuracy.

What to watch for when evaluating AI benchmarks

• Who ran the benchmark. A vendor-run benchmark is a marketing claim; an evaluation executed by an independent third party is verification.
• Reasoning effort. Meta ran tests on “max” reasoning; default API and consumer app settings operate at lower compute budgets.
• Model variant. Wonder Full achieved the highest scores, but the public API only serves standard Wonder.
• Scoring methods. Partial credit, binary pass/fail, and composite index ratings reflect completely different performance realities.
• Real-world workloads. Across my own 23 practical tasks, all five flagship models delivered 99-100% success rates - the true differences came down to latency and API costs.

Frequently asked questions

Does the Personal Superintelligence Index measure true superintelligence?

No. It is a proprietary composite of seven personal and administrative office tasks selected by Meta. The label is branding.

How many benchmarks did Muse Wonder actually win?

Muse Wonder Full leads in 7 out of the 14 rows shown on Meta’s table. Standard Muse Wonder, the model actually available via API, outperforms Opus 5.5 in 5 rows.

Which benchmarks are independently verified?

Harvey LAB-AA, AutomationBench-AA, Terminal-Bench 4.0-AA, and the AA Index are measured by Artificial Analysis; Vals Finance by Vals AI; GDP.pdf by Surge; DeepSWE by Datacurve; and SWE-Marathon by Abundant AI. PSI, MemoryBench, and internal components like FocusBench and VBT were run internally by Meta.

Where does Muse Wonder fall behind?

Based on Meta’s own published charts: ProgramBench (59.0 versus 70.5 for Opus), HealthBench Pro, Terminal-Bench 4.0, OSWorld, and GDP.pdf.

// read moreMeta Muse Wonder: Benchmarks, API Pricing, and Full TestBenchmarksMuse Wonder vs Opus 5.5 vs GPT-6 Astra: 23 Real-World TestsBenchmarksTop Muse Wonder Alternatives for Frontier AI TasksAI Tools

Kinetyra