On October 12, Meta officially announced the Muse Wonder and Muse Wonder Full model family, positioning them as frontier models designed for agentic tasks across professional work and personal life. The announcement comes just six weeks after Meta released Muse Spark 1.3, marking a shift from lightweight, fast-response models toward large-scale models built to handle complex multi-step workflows.

Currently, Muse Wonder is available to developers via the Meta Model API and the Muse Code platform, while the flagship Muse Wonder Full is exclusively available to Muse Code subscribers. Meta noted that features on the personal Muse app will be rolled out gradually to paid subscribers.

Active Memory and Active Focus Target Multi-Step Tasks

Meta identified context retention and persistence through failures as the two biggest hurdles facing current AI agents. To address this, the company introduced two core capabilities: Active Memory and Active Focus.

Active Memory is tasked with carrying critical user context across different conversation sessions while proactively retrieving relevant details as new tasks arise. Meanwhile, Active Focus helps the model sustain effort toward its objective when encountering challenges, automatically testing alternative solutions, working around bottlenecks, and asking the user for clarification before executing high-stakes actions.

In software development, these two capabilities enable agents to adhere to the requirements of an entire codebase, carrying out modifications across multiple test-and-debug cycles. The company stated that the models were trained to pursue long-horizon tasks and had their focus evaluated using benchmarks such as FocusBench and SWE-Marathon.

A Leap Forward in Agentic Capabilities

According to an independent report by Artificial Analysis, on the Intelligence Index v4.3, Muse Wonder Full scored 56 points and Muse Wonder scored 52 points at maximum reasoning effort. Both results significantly surpass the 48-point score of Muse Spark 1.3, which was released last month.

Muse Wonder Full placed among the top three models in six out of ten tests on the Intelligence Index. The model led in the CritPt scientific reasoning benchmark at 33% and the GDP.pdf document test at 33%, while ranking second on the GDPval-AA v2.1 leaderboard with an Elo rating of 1,857, placing it within the confidence interval of Claude Opus 5.5 at 1,866 Elo and Claude Sonnet 5.5 at 1,838 Elo.

In the coding agent category on Muse Code, Wonder Full achieved 68 points at maximum reasoning effort, matching Sonnet 5.5 on Claude Code and surpassing Opus 5.5’s 66 points.

Additionally, Artificial Analysis recorded Wonder Full’s hallucination rate dropping to 23%, the second-lowest among all models scoring 45 or higher on the intelligence index, trailing only Gemini 4 Argon at 15%. However, general knowledge on the AA-Omniscience benchmark remains a relative weakness, with Wonder Full scoring just 30 points and Wonder scoring 25 points, trailing competitors like Opus 5.5 at 46 points and GPT-6 Astra at 43 points.

Task Completion Costs Surge with Input Token Volume

Both new models retain a 1-million-token context window, support multimodal inputs including text, images, video, and audio, and produce text-only output. Output token consumption for both models is approximately 54,000 tokens per task, down roughly 10% compared to their predecessor.

However, pricing and large input volumes drive up real-world costs. The Muse Wonder API is priced at $4 per 1 million input tokens, $0.20 for cached tokens, and $20 per 1 million output tokens. This pricing matches Claude Opus 5.5 and is double that of models like Claude Sonnet 5.5, Gemini 4 Argon, and GPT-6.1 Sol, which cost $2 for input and $10 for output.

Because each task consumes an average of 7.69 million input tokens, the average cost to complete a task with Muse Wonder reaches $4.52, 2.8 times higher than Muse Spark 1.3 at $1.60. Meanwhile, Muse Wonder Full is not sold on a per-token basis and is only available through a Muse Code subscription.

The debut of the Wonder family highlights Meta’s push to compete directly with flagship models from Anthropic and OpenAI on complex agentic tasks rather than competing solely on price. However, with token pricing and massive context ingestion driving per-run costs higher, the next question is how developers will balance budgets before integrating these agents into production automation workflows.