Meta Enters Frontier AI Race with Muse Wonder Launch

Meta has officially entered the top tier of the artificial intelligence race with the release of its Muse Wonder and Muse Wonder Full models. Unveiled on Monday, these systems represent the first major flagship offering from Meta Superintelligence Labs designed to compete directly with leading software from OpenAI and Anthropic.

Chief Executive Officer Mark Zuckerberg announced the rollout on Threads, describing the models as a major leap toward personal superintelligence. Interestingly, the terminology aligns closely with recent remarks from President Trump, who asserted that the White House favors the term 'Super Intelligence' over standard industry phrasing. Meta rebranded its advanced AI division in 2025, positioning the company neatly within the political nomenclature.

Performance and Benchmarks

The launch follows a turbulent period for Meta's AI ambitions. Following mixed receptions and performance discrepancies with its open-source Llama 4 release in April 2025, Zuckerberg restructured the division. He orchestrated a roughly $14 billion deal for Scale AI and appointed its founder, Alexandr Wang, to lead the new Superintelligence Labs. The strategy pivot marked a definitive shift away from open-source models toward closed systems.

Meta's latest models introduce two core capabilities: Active Memory, which retains context across separate conversations, and Active Focus, designed to maintain progress on complex goals while consulting users at critical junctures. According to Meta's internal Personal Superintelligence Index, the Wonder Full variant scored 59.1, outperforming Anthropic's Claude Opus 5.5 at 55.2 and OpenAI's GPT-6 Astra at 54.5.

In specialized evaluations, such as the SWE-Marathon coding test and Harvey's Legal Agent Benchmark, Meta's models posted strong numbers. However, the company noted that certain tests, including the coding marathon, ran at maximum effort without strictly controlled comparisons. Furthermore, on MemoryBench, a test tracking factual recall across long conversations, both Wonder models scored approximately 85.5, edging out competing systems from Anthropic and OpenAI.

Independent benchmarks present a more measured view. Artificial Analysis, which compiles weighted evaluations across ten criteria, ranks Anthropic's Opus 5.5 first with a score of 58. Wonder Full ties with Anthropic's cheaper Claude Sonnet 5.5 at 56, while the commercial Muse Wonder developer version scores 52.

Why it matters

Beyond raw performance, Meta is undercutting rivals on cost while introducing aggressive data-gathering features. Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens—where tokens are the fundamental word fragments processed by AI. This matches Anthropic's pricing for Opus 5.5 and sits significantly below OpenAI's $10 input and $50 output fees for Astra.

Yet, the emphasis on persistent memory raises substantial privacy questions. Critics remain wary of Meta's data handling history, particularly following recent disclosures regarding unintended message synchronization and prior sandbox security escapes during internal testing.

What to watch

Observers will monitor independent developer adoption to see whether Meta's competitive pricing and memory capabilities translate into real-world market share. Additionally, scrutiny will focus on how the company manages data privacy and model alignment as enterprise customers test the limits of active memory agents.