Meta launches Muse Wonder AI models

On SWE-Marathon, Meta's own published figures credit Wonder Full with a score of 79.0 ^[1], essentially matching or beating OpenAI's GPT-6 Astra (74.1) and Anthropic's Claude Opus 5.5 (74.2) on paper. But when Artificial Analysis ran its own independent testing, Wonder Full scored only 57.5 on that same benchmark ^[2]- still ahead of Opus 5.5's 46.3 and GPT-6 Astra's 42.5 by that measure ^[2], but more than 20 points below what Meta told the public. That gap sat at the center of the day's skepticism: independent testers measured something meaningfully different from what Meta reported, and the discrepancy raises questions about benchmark methodology industry-wide, not just about Meta ^[2].

Elsewhere the independent numbers were kinder to Meta. On Artificial Analysis' Intelligence Index, Wonder Full scored 56, tying Claude Sonnet 5.5 and landing just two points behind Opus 5.5's 58 ^[3]- a genuinely frontier-tier result. The standard Wonder model scored 52 ^[2]. Meta also highlighted a safety figure: on an agentic misalignment test checking whether a model will act harmfully to avoid being shut down or replaced, Wonder Full logged a 0% harmful-action rate ^[4]. Reaction split along platform lines: Meta's own launch announcement and an enthusiastic insider post from a Meta researcher struck a triumphant note, while Reddit threads stayed skeptical of Meta's self-reported benchmark claims generally, with one commenter pointing to a task-level breakdown on a coding benchmark that looked far less favorable to Wonder than the headline numbers suggested.

Wonder's launch lands as Meta's consumer Muse personal-agent app approaches 9 million downloads since its September 2026 debut, but growth is decelerating and retention trails ChatGPT ^[5][6]. That context matters: Wonder and Wonder Full are not just a research flex, they're the engine meant to make Muse's paid tiers worth the money - Power at $20 a month for 500 million tokens a week, Maximum at $100 a month for 3 billion tokens a week, against a free tier capped at 100 million tokens a week ^[7].

On the developer side, Meta priced API access to Wonder at $4 per million input tokens and $20 per million output tokens, with cached input discounted to $0.20 per million ^[2]- pricing that lines up closely with what Anthropic charges for Opus 5.5, a parallel that developer communities picked up on quickly as a sign Meta is matching incumbent pricing rather than undercutting to win share. Paired with the API-versus-subscription split that gates the strongest model, the pricing structure makes Wonder as much a monetization experiment for the Muse product line as it is a model release.

Wonder's headline features are Active Memory, which carries and retrieves important context about a user across conversations, and Active Focus, which keeps the model working through a hard goal by trying alternative approaches and adapting when it hits a blocker ^[8]. Both are explicitly built for the kind of sustained, personal-assistant use case Meta wants Muse to own.

That ambition runs straight into a problem Meta was already managing: its Muse personal agent faced a privacy incident in which it reportedly synced users' private iMessages without consent ^[2], and the broader rollout of persistent-memory agent features had already drawn privacy and safety scrutiny before Wonder shipped ^[9]. Giving the model an explicit, named capability for remembering personal context across days, weeks, and months raises the stakes on exactly the behavior that caused the earlier backlash, and it's a tension Meta's launch materials don't address directly. A broadcast panel discussion of Meta's AI push put the question bluntly - debating whether it amounts to "inspiration or intrusion" - and noted that Meta is pressing ahead with its AI ambitions even as rivals have shown more public caution about agentic-AI risk.

Wonder is the product of a fast, well-funded pivot. Meta founded Meta Superintelligence Labs in the summer of 2025, installing Alexandr Wang as Chief AI Officer alongside a $14.3 billion investment for a 49% non-voting stake in Scale AI ^[10]. The lab's first model, Muse Spark, arrived in April 2026, scoring 52 on the Intelligence Index and marking a deliberate break from Meta's prior open-source Llama strategy toward a closed, frontier-style model tier ^[11][12]. Wonder and Wonder Full, released October 12, 2026 and trained at Meta's new gigawatt-scale Prometheus data center in Ohio ^[13], are the first Meta models put head-to-head against OpenAI's and Anthropic's current flagships ^[14].

Meta's own framing leans hard into parity with the leaders. Chief AI Officer Alexandr Wang called Wonder the best model in the world for personal and professional agents ^[15], and Meta research scientist Kavya Kopparapu said the team put tremendous effort into making the model remember what matters to a user over days, weeks, and months ^[13]. It's the first time Meta has put a model directly on the same benchmark charts as OpenAI's and Anthropic's flagships ^[16], even though Opus 5.5 still leads the Intelligence Index outright.