Meta introduces Muse Wonder and Muse Wonder Full AI models
Meta unveiled two new artificial intelligence models on October 12, 2026—Muse Wonder and Muse Wonder Full—positioning them as advanced systems designed for autonomous agent work across professional domains. The company describes the models as built for tasks including software development, financial analysis, legal document review, and workflows requiring coordination across multiple steps.
The models represent a shift in how AI systems approach complex assignments. Rather than simply responding to individual prompts, agentic AI systems like Muse Wonder autonomously plan multi-step workflows, use tools and feedback loops, and pursue specific goals with minimal human oversight. Meta's implementation centers on two core features: Active Memory, which retains relevant information across extended conversations, and Active Focus, which enables the models to persist toward objectives, attempt alternative approaches after failures, and request confirmation before taking significant actions.
According to Meta, both models process text, images, audio, and video inputs while generating text outputs, with context windows accommodating approximately 1 million tokens. This positions them competitively in a market where context window capabilities have become a key differentiator, though pricing structures for such large contexts vary significantly across providers.
Benchmark performance and competitive positioning
Independent testing by Artificial Analysis showed measurable improvements over Meta's previous generation. On version 4.3 of the Intelligence Index—a weighted average of 10 production benchmarks across agents, coding, scientific reasoning, and general capability—Muse Wonder scored 52 and Muse Wonder Full reached 56 at maximum reasoning effort. Both surpassed the 48 score achieved by Meta's earlier Muse Spark 1.3 model. The Intelligence Index methodology assigns particular weight to agentic tasks at 30 percent, followed by coding and scientific reasoning at 20 percent each, and general capability at 30 percent.
The largest performance gains appeared in evaluations focused on agent capabilities, though benchmark results cannot guarantee consistent performance across the full range of real-world applications users may attempt.
Pricing and availability
Meta made Muse Wonder available through the Meta Model API and Muse Code, pricing it at $4 per million input tokens, $20 per million output tokens, and $0.20 per million cached input tokens. This represents a significant increase from Muse Spark 1.3, which launched on September 2, 2026 at $1.25 per million input tokens and $4.25 per million output tokens.
However, the pricing undercuts major competitors. OpenAI's GPT-5.5 costs $5 per million input tokens and $30 per million output tokens, while Anthropic's Claude Opus models are priced at $5 and $25 per million tokens respectively. Meta's lower pricing positions Muse Wonder as a cost-competitive alternative in the frontier model market.
Muse Wonder Full, which achieved the higher benchmark score, remains limited to Muse Code subscribers and does not offer public pay-as-you-go API access. Meta indicated that Muse Wonder will eventually power features in its Muse assistant, beginning with subscriber accounts.
Strategic departure from open weights
The launch marks a continuation of Meta's strategic shift away from open-weight model releases. The company previously built its AI reputation on the Llama series, with Llama 3.1 405B—released in 2024 and described at the time as the largest open-source AI model in history—representing the pinnacle of that approach. Meta transitioned to proprietary closed models with the Muse series beginning in April 2026, moving away from the downloadable, open-weight philosophy that characterized its earlier work.
Neither Muse Wonder nor Muse Wonder Full offers open-weight downloads, continuing the closed approach established with the Muse line. This represents a fundamental business model change from the company's previous strategy of releasing weights publicly.
Meta stated that the models incorporate safeguards against privacy violations, cyber misuse, and harmful requests, while noting that comprehensive safety documentation remains forthcoming. Whether the benchmark improvements translate into reliable autonomous performance beyond controlled demonstrations and evaluations remains an open question as users begin testing the models in production environments.
