Meta Unveils Muse Wonder Models for Complex Agent Tasks
Meta has introduced Muse Wonder and Muse Wonder Full, two new AI models designed to handle intricate multi-step assignments through autonomous agents. The company positions these models as its most capable yet for tackling extended workflows in professional and personal contexts.
Meta Platforms Inc. unveiled Muse Wonder and Muse Wonder Full today, positioning both as its strongest artificial intelligence models for powering agents that execute personal and professional workflows. The company engineered these systems to tackle extended, complex assignments that users typically delegate to AI—such as navigating insurance claims. Meta highlights particular strengths in code generation alongside specialized domains including legal work, financial services and healthcare.
Two capabilities distinguish the new models. Active Memory preserves user context across separate conversations, surfacing relevant information when needed. Active Focus maintains model performance during challenging tasks by attempting alternative strategies, recovering from mistakes and consulting users before executing critical actions.
Meta developed both models using its Prometheus data center operating at gigawatt scale, extending techniques from Muse Spark, which the company launched in April. The models accept multimodal input combining text with video, audio and images within a 1 million-token context window, though responses are delivered in text only.
Benchmark Performance
Meta's model card compares Muse Wonder Full against OpenAI Group PBC's GPT-6 Astra and Anthropic PBC's Claude Fable 5.1 and Claude Opus 5.5. Most evaluations ran competing models at maximum reasoning effort, though Meta notes its own runs of rival systems were best-effort and prompts and tools may not have been optimally configured for them.
On Artificial Analysis' adaptation of Harvey's Legal Agent Benchmark, Muse Wonder Full achieved a 13.1% pass rate on tasks requiring complete criterion satisfaction and hallucination verification, compared with 8.6% for GPT-6 Astra.
Coding benchmarks showed mixed results. Muse Wonder Full led on two extended-horizon tests: DeepSWE 1.1 and SWE-Marathon, scoring 57.5% on the latter versus 46.3% for Claude Opus 5.5. The model card notes SWE-Marathon results lack controlled comparison status. On ProgramBench, which requires agents to reconstruct entire programs from scratch, Claude Opus 5.5 prevailed at 70.5% to 59%. Claude Opus 5.5 also outperformed on HealthBench Pro and OSWorld V2.1, and scored two points higher on Artificial Analysis Intelligence Index at 58.
Safety and Risk Assessment
Meta plans to release a comprehensive safety report shortly. The model card indicates Muse Wonder Full may reach high-risk thresholds for chemical and biological threats when operating without safeguards. When safety post-training and system-level classifiers are active, the model rates "moderate or lower" across all three risk domains in the Meta Superintelligence Scaling Framework, encompassing cybersecurity and loss of control.
On ExploitGym's cyber benchmark, Muse Wonder Full identified 92 flags, with verification confirming 21 exploited the intended vulnerabilities and 71 leveraged unintended weaknesses. Since ExploitGym's published rankings count only intended routes, the company emphasizes the 21 figure for comparison. This disclosure follows Meta's revelation two months earlier that one of its models breached an external organization during a cybersecurity test due to a sandbox configuration error.
Meta's research describes Muse Wonder as "our most aligned model to date." On an agentic misalignment test measuring whether agents perform harmful actions to achieve objectives or avoid replacement and shutdown, Muse Wonder Full registered a 0% harmful-action rate.
Availability and Integration
Muse Wonder replaces Muse Spark 1.3 at the top of Meta's model lineup. Muse Spark 1.3, released by Meta Superintelligence Labs six weeks prior, was described by Chief AI Officer Alexandr Wang to Bloomberg as "competitive" with Claude Fable 5.1. Meta's Muse personal agent has operated on Muse Spark since its U.S. debut on September 8.
Developers can access Muse Wonder immediately through the Meta Model API and Muse Code, Meta's coding agent introduced in August. Muse Wonder Full availability is restricted to Muse Code subscribers. Meta intends to power its Muse agent with Muse Wonder soon, with subscribers receiving access first.
Read more
Speed to operational status now matters more than GPU counts in AI infrastructure buildouts
As organizations race to deploy AI factories, the bottleneck has shifted from securing graphics processors to managing data pipelines and storage. Infrastructure vendors are rethinking how they approach these projects from the ground up.
SpaceXSI's Grok Voice 3.0 Tops Industry Benchmarks in Speech-to-Speech Showdown
Elon Musk's SpaceXSI has unveiled Grok Voice 3.0, a speech-to-speech model that outperforms competitors from OpenAI and Google across multiple independent testing frameworks.
Microsoft Pursues Composable Memory to Ease AI Data Center Bottlenecks
As memory scarcity constrains AI workloads across data centers, Microsoft is exploring composable memory architectures that dynamically allocate capacity across servers rather than locking it within individual machines.