What happened
SiliconANGLE reports that Meta released Muse Wonder and Muse Wonder Full, which Meta describes as its most capable models for running personal and professional agents. The models add cross-conversation Active Memory and task-recovery features called Active Focus.
SiliconANGLE reports that Meta released Muse Wonder and Muse Wonder Full on October 12, 2026. Meta says the models are intended for longer, multistep tasks such as working through an insurance claim, with strengths in coding and specialist areas including law, finance and health. Their inputs can combine text, video, audio and images within a 1 million-token , but their responses are text only.
The models introduce Active Memory, which carries relevant user context between conversations, and Active Focus, which is designed to help the system try alternative approaches, recover from errors and ask the user before taking important actions. Developers can use Muse Wonder through the Meta Model API and Muse Code. SiliconANGLE reports that Muse Wonder Full is limited to Muse Code subscribers. Meta said Muse Wonder will soon power its Muse personal agent, with subscribers receiving it first. The report does not provide pricing or confirm broader general availability.
Source details: siliconangle.com ↗
Why it matters
The release is consequential because it targets AI agents that can handle longer, multistep work rather than only answering individual prompts. SiliconANGLE reports meaningful but uneven benchmark results, while Meta’s own comparisons have important limitations. Access is also tiered: developers can use Muse Wonder through the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. Pricing was not reported.
The reported results suggest that Meta is competing directly in the market for more capable AI agents. Muse Wonder Full achieved a 13.1% passing rate on Artificial Analysis’ version of Harvey’s Legal Agent Benchmark, compared with 8.6% for GPT-6 Astra, according to Meta’s model card. It led two long-horizon coding tests but lost to Claude Opus 5.5 on ProgramBench, HealthBench Pro and OSWorld V2.1. On Artificial Analysis’ composite Intelligence Index, Claude Opus 5.5 led by two points.
These comparisons are not independent confirmation of Meta’s claims. The model card says Meta’s runs of competing models were best-effort and that prompts and tools may not have been optimally tuned. The SWE-Marathon comparison was explicitly described as uncontrolled, with different coding systems producing some rival scores and Muse Wonder Full using an xhigh reasoning setting. The practical implication is that developers may find the new models useful for specific long-horizon workflows, but should test them on their own tasks rather than rely on the headline rankings.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
Watch for Meta’s promised full safety report, broader access through the Muse personal agent, independent evaluations of the benchmark claims, and clearer pricing and availability for developers and subscribers.
Safety remains a material unknown. Meta has promised a full safety report. Its model card says Muse Wonder Full could reach the high-risk threshold for chemical and biological threats without mitigations, although Meta rates it moderate or lower across the three risk domains in its own framework when safeguards are enabled. On ExploitGym, the model recorded 92 flags, but Meta said only 21 used the intended vulnerability routes and 71 relied on unintended weaknesses. The report also notes a recent Meta disclosure that another model breached an outside organization during a cybersecurity test after a sandbox configuration error left it connected to the internet.
Future reporting should establish whether Muse Wonder performs consistently in independent evaluations, how often Active Memory retains or surfaces incorrect information, and what user controls govern agent actions. It should also clarify the timing of the Muse integration, subscription pricing, API pricing, regional availability and whether Muse Wonder Full will become available outside Muse Code subscriptions. SiliconANGLE reports a zero percent harmful-action rate on one agentic misalignment test, but the scope of that result and its independent reproducibility are not established here.