Agents

Meta Launches Muse Wonder, Its First Frontier AI Model

Oct 13, 2026·5 min read·By Rahish Saifi

On October 12, 2026 (US time), Meta introduced Muse Wonder and Muse Wonder Full—a pair of frontier models the company positions as its most capable release yet for personal and professional AI agents. In a research blog post dated that day, Meta described the series as the next step on its scaling ladder after Muse Spark, trained at the gigawatt-scale Prometheus data center. For developers and subscribers, the launch is not a distant roadmap item: Muse Wonder is available through the Meta Model API and Muse Code now, while Muse Wonder Full is available today for Muse Code subscribers.

What Meta launched

Meta frames Muse Wonder around two agent-centric ideas. Active Memory is the label for carrying important user context across conversations and retrieving details when they matter again. Active Focus describes sustained work toward a goal—trying alternate approaches, working through blockers, asking for clarification, and checking in before consequential actions. The blog ties those themes to demos such as multi-step planning, coding across a codebase, and multimodal tasks (for example, building a playable 3D kitchen game from image inputs).

According to Meta’s model card, both variants accept text, image, video, and audio as input and produce text output, with a 1 million-token context window at inference. Intended uses span general reasoning, software engineering, personal and professional agents, domain work (law, finance, health), multimodal understanding, and R&D support. Meta notes that Muse Wonder will soon power experiences in the Muse app, rolling out first to subscribers, and will also drive artifact generation and other higher-intelligence processes.

Reporting from The Business Times (Bloomberg, Oct 13, 2026) adds that the model was known internally by the codename “Watermelon.” Meta AI chief Alexandr Wang told the outlet that Wonder can carry user information across “days, weeks, months” so people do not have to repeat themselves, and that the agent can try new tactics when it hits dead ends.

How Artificial Analysis scores the release

Meta’s blog highlights benchmark tables comparing Wonder models with other frontier systems; for an independent read, Artificial Analysis published its own take on October 12, 2026. The firm reports Artificial Analysis Intelligence Index v4.3 scores at max reasoning effort of 52 for Muse Wonder and 56 for Muse Wonder Full—versus 48 for Muse Spark 1.3 (max)—a gap Artificial Analysis characterizes as a material jump only six weeks after Spark. Those figures are the analyst’s measurements, not a universal ground truth; leaderboard math can differ from product experience.

Artificial Analysis also claims Muse Wonder Full (max) places in the top three on six of the ten Intelligence Index evaluations, with particular emphasis on agentic knowledge work and scientific reasoning in its write-up. On coding agents, it reports Muse Wonder Full at 68 on its Coding Agents Index in Muse Code (alongside cited comparisons to Claude Code configurations), and Muse Wonder (max) at 63, up from 54 for Muse Code with Muse Spark 1.3 (max). Meta’s model card similarly cites Artificial Analysis for several agent benchmarks (for example Harvey LAB-AA and AutomationBench-AA) while mixing in provider and third-party scores elsewhere—worth remembering when comparing rows.

Coverage via The Economic Times (AFP, Oct 13, 2026) notes that Meta pointed to Artificial Analysis data that appears to show Wonder outperforming Spark as well as various models from Anthropic, OpenAI, and Google—while also flagging that Wonder ranks among the more expensive models on Artificial Analysis’s site.

Pricing and availability

For developers weighing API economics, Artificial Analysis lists Muse Wonder token pricing at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens. It states that Muse Wonder Full is not priced per token and that there is no public “cost per Intelligence Index task” figure for Full because of that. Artificial Analysis also estimates Muse Wonder (max) at about $4.52 per Intelligence Index task—higher than Muse Spark 1.3 (max) at $1.60 in its methodology, driven by both higher list prices and higher input-token use per task in its runs.

Artificial Analysis lists Muse Wonder Full as a proprietary model (weights not open). Availability at launch: Wonder on Meta’s first-party API and Muse Code; Wonder Full only for Muse Code users, per Artificial Analysis and Meta’s blog. The Business Times reports Wang saying Wonder is larger and more expensive to operate than Muse Spark, and that Meta is already developing specialized versions aimed at retaining much of Wonder’s capability at a cost suitable for billions of users.

Safety and governance

Frontier agents that read messages, calendars, and documents raise the stakes on privacy and misuse. Meta’s blog says Muse Wonder offers its strongest safeguards to date against CBRN and cyber misuse while preserving defensive cyber capabilities under the Meta Superintelligence Scaling Framework, and that a full safety report will be available soon.

The model card’s risk section (assessed on Muse Wonder Full) states that without mitigations Meta does not rule out the model reaching a high-risk threshold in the chemical and biological domain, so it applies mitigations including safety post-training and system-level classifiers. For cybersecurity and loss of control, Meta reports evaluations place the unmitigated model at “moderate or lower” before mitigations, and the same level across domains once mitigations and safeguards are in place. That language describes Meta’s internal framework—not a public regulator sign-off.

Why it matters and what’s next

For Meta, Wonder is a credibility test after the company moved away from the open Llama line toward the closed Muse Spark family in 2026. CEO Mark Zuckerberg, quoted on X in AFP’s Economic Times piece, described Muse Wonder as “a much larger and more advanced class of models that brings us to the frontier” after Spark. Wang’s posts and interview remarks—reported by AFP and Bloomberg via Business Times—claim parity with leading labs on agentic work, a narrative Meta needs if Wonder is to power a paid Muse subscription agent at scale.

For builders, the practical split is clear: Wonder is the API-facing frontier model with published per-token pricing on Artificial Analysis’s tracker; Wonder Full is the heavier subscriber tier inside Muse Code without a public token meter. Meta says Wonder will eventually flow into consumer Muse features and into more specialized models derived from Wonder—while Wang told Business Times that Meta Superintelligence Labs is already training “larger and more scaled models” beyond this release.

Independent verification still comes down to hands-on use: agent harnesses, tool policies, and memory stores can dominate outcomes long after benchmark points fade. If Wonder’s Active Memory and Active Focus work as advertised in real accounts—not just in launch demos—they could matter more than a few Intelligence Index points for the personal-agent race Meta is explicitly prioritizing over a coding-only strategy.