The new Muse Wonder and Wonder Full models introduce enhanced memory and agentic capabilities, with Meta reporting competitive performance against leading AI systems.
Meta has introduced two new artificial intelligence models, Muse Wonder and Muse Wonder Full, as the company expands its efforts to compete with advanced AI systems developed by OpenAI, Anthropic and Google.
Released on October 12, 2026, the models were developed by Meta Superintelligence Labs and are designed to support complex reasoning, software development and multi-step tasks performed by AI agents.
The launch marks a further expansion of Meta's Muse model family, following the introduction of earlier systems focused on speed and everyday AI assistance.
According to Meta, Muse Wonder and Wonder Full incorporate capabilities called Active Memory and Active Focus, intended to help AI systems retain relevant information and maintain attention across longer tasks.
These features are designed to support applications requiring multiple stages of reasoning and execution, including coding, document analysis and professional workflows.
Meta said the models demonstrated competitive performance against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 in selected benchmark evaluations.
However, performance varies across tests, and the results do not establish that Meta's models outperform competing systems across all tasks.
On the Artificial Analysis Intelligence Index, Muse Wonder Full recorded a score of 56, compared with 52 for Muse Wonder. Meta's earlier Muse Spark 1.3 model scored 48.
The company also reported improvements in agentic performance, including software engineering and tasks requiring sustained interaction with digital tools.
Both models support a context window of one million tokens, allowing them to process substantial amounts of information within a single interaction.
Muse Wonder is available through the Meta Model API and Muse Code, while the more capable Wonder Full model is initially restricted to Muse Code subscribers.
API pricing for Muse Wonder has been set at $4 per million input tokens and $20 per million output tokens.
Meta plans to integrate the new models into its Muse AI assistant, initially for subscribers, before expanding availability more broadly.
The company has positioned the models as part of its longer-term strategy to develop AI assistants capable of remembering context, handling complex requests and completing tasks on behalf of users.
Such capabilities could support enterprise applications involving customer service, business operations, software development and information management, although their effectiveness will depend on implementation and reliability.
The launch follows Meta's continued investment in AI infrastructure, research and talent as competition among major technology companies intensifies.
Meta Chief Executive Mark Zuckerberg has described the company's broader objective as developing personal superintelligence, with AI systems becoming more capable of assisting users across everyday activities.
While the reported benchmark results indicate progress, independent evaluations and real-world deployments will provide further evidence of how the models perform against established competitors.
The announcement places Meta's latest models within an increasingly competitive market for advanced AI systems, particularly those designed to support autonomous and agent-based workflows.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.