Model

Muse Wonder

CLOSED

Meta · Muse Wonder family · released Oct 12, 2026

Meta's first frontier-tier model, released October 12, 2026 in the Meta Model API (muse-wonder-1.0) and Muse Code at $4/$20 per Mtok ($0.20 cached input), four times Muse Spark 1.3's price. Meta's own scorecard puts it level with Claude Opus 5.5 and GPT-6 Astra rather than ahead: it leads DeepSWE 1.1 (78.8) and Meta's internal MemoryBench, trails Opus 5.5 on HealthBench Pro, OSWorld and Terminal-Bench 4.0. Independent rows here come from Vals AI's launch-day run (Vals Index 61.06%, #10 of 47; #1 on CyberBench v1.1). NOT FILED: Meta's Personal Superintelligence Index (54.3) and MemoryBench (85.6) are internal composites; Harvey LAB-AA (10.0) is Artificial Analysis's v1.1 re-grading with hallucination checks, not the Vals-run harvey-legal-agent column; AutomationBench-AA (39.1) is AA's variant, not Zapier's private-set AutomationBench column; ProgramBench (50.0) uses a >=95%-tests-passing metric that disagrees with Vals' ProgramBench scale; Artificial Analysis Intelligence Index v4.3 is 52 (no corpus column).

ReasoningCodingVisionFunction callingTool useAgentic

3017.3

Elo · rank #5

Parameters

Undisclosed

Active params

—

Context

1M tokens

Architecture

Proprietary multimodal reasoning model (architecture undisclosed); 1M-token context, trained for long-running personal and professional agent work

License

Proprietary

Languages

—

API price (in/out)

$4 / $20

Modalities

text · vision · audio · video

Benchmark results

Bar shows position within the tracked field; marker = field best

Arena EloHuman preference1513#1

best: this model · 1513

DeepSWECoding78.8%#2

best: Muse Wonder Full · 79.0%

GDP.PDFVision30.8%#6

best: GPT-5.6 Sol · 40.0%

Harvey Legal Agent BenchmarkAgents17.1%#5

best: Muse Spark 1.2 · 25.4%

HealthBench ProfessionalKnowledge60.5%#8

best: Claude Opus 5.5 · 65.6%

JobBenchAgents70.5%#2

best: Muse Wonder Full · 75.3%

OSWorld 2.0Agents72.6%#5

best: Claude Opus 5.5 · 81.8%

SWE-MarathonCoding53.1%#2

best: Muse Wonder Full · 57.5%

Terminal-Bench 4.0Agents49.5%#12

best: Claude Sonnet 5.5 · 70.6%

Terminal-Bench Science 0.1Agents22.9%#10

best: GPT-6 Astra · 62.9%

Vals Finance AgentAgents56.6%#16

best: Gemini 4 Argon · 65.4%

Run it locally

Closed weights — available via API only. No local deployment.

Input / M tok

Output / M tok

$20

Muse Wonder family

Elo progression across releases

Muse WonderOct 20263017.3 Muse Wonder FullOct 20263100.5

Resources

DOCShttps://research.meta.ai/blog/muse-wonder APIAvailable via Meta API

API price $4/$20 · each benchmark row carries its own source badge (see methodology)