The company introduced Muse Wonder and Muse Wonder Full on October 12, describing them as its most capable models, designed for complex professional assignments, software development and autonomous digital assistants. Chief executive Mark Zuckerberg said the release demonstrated that Meta had reached the frontier of artificial intelligence development.
The models emerged from Meta Superintelligence Labs following an extensive restructuring of the company’s AI operations. Their performance represents a substantial improvement over Muse Spark, the smaller model introduced earlier this year, particularly in coding, financial analysis, legal reasoning and tasks requiring sustained execution.
Independent evaluations have provided support for Meta’s performance claims, although results vary considerably across individual tests. Artificial Analysis assigned Muse Wonder Full a score of 56 on its Intelligence Index, compared with 52 for Muse Wonder, 53 for OpenAI’s GPT-6 Astra and 58 for Anthropic’s Claude Opus 5.5.
The results place Meta’s strongest model within the competitive range of established frontier systems without establishing an overall lead. Performance differences across specialised benchmarks also indicate that individual models retain advantages depending on the complexity and nature of assignments.
Meta’s published evaluations show Muse Wonder Full achieving a score of 79.0 on the DeepSWE software engineering benchmark, compared with 74.1 for GPT-6 Astra and 74.2 for Claude Opus 5.5. On SWE-Marathon, which measures performance on extended engineering assignments, Wonder Full scored 57.5 against 42.5 and 46.3 respectively.
However, the model trailed competing systems on some other assessments. GPT-6 Astra scored 59.1 on Terminal-Bench 4.0, compared with 57.1 for Wonder Full, while Anthropic’s model performed better on certain professional tasks.
The distinction matters because Meta is positioning the models primarily as engines for autonomous agents rather than conventional conversational assistants. Such systems are designed to complete complicated assignments involving multiple applications, documents and decisions, with limited intervention from users.
Two capabilities, called Active Memory and Active Focus, form central elements of the new architecture. Active Memory enables the models to retain relevant information across conversations, while Active Focus allows them to pursue objectives through successive steps, recover from errors and seek clarification when necessary.
Meta said these features could support activities ranging from interpreting insurance claims and rental agreements to analysing financial information and developing software applications. The company also highlighted potential applications involving healthcare questions, business workflows and computer operation.
Muse Wonder is available through the Meta Model API and Muse Code, the company’s programming environment. The more powerful Muse Wonder Full is restricted to subscribers using Muse Code, creating different access arrangements for developers and commercial customers.
Meta has priced the standard model at $4 per million input tokens and $20 per million output tokens. Cached input costs $0.20 per million tokens. Both versions support context windows of approximately one million tokens, allowing substantial quantities of information to be processed within individual assignments.
The company offers Muse Code subscriptions that include access to Wonder Full, with higher-tier plans priced at $50 and $200 monthly. These arrangements establish a commercial distribution model alongside Meta’s existing consumer AI services.
Muse Wonder will also become available within the Muse personal assistant, initially for subscribers. Meta expects the integration to support more demanding tasks that require continuity between conversations and interactions with external services.
The launch follows substantial spending on computing infrastructure, specialist recruitment and research under Zuckerberg’s strategy of developing personal superintelligence. Meta has sought to strengthen its competitive position after earlier model releases failed to establish consistent leadership across industry benchmarks.
The company said its models were trained to protect private information, reject harmful requests and assist legitimate cybersecurity work. It acknowledged that personal agents may need access to sensitive material, including messages, calendars and documents.
Meta said its development and deployment procedures follow the company’s Superintelligence Scaling Framework, which establishes standards for evaluating advanced AI capabilities and associated risks. A fuller safety evaluation is expected to be published separately.
Developer documentation specifies that information submitted through the standard operating mode is not used for model training. A separate contributor mode allows subscribers to exchange permission for training use of their content for higher usage allowances, with an option to disable that setting.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
