Meta Goes Frontier With Muse Wonder, but the Scoreboard Still Shows Claude on Top
On Monday, Meta's Superintelligence Lab shipped Muse Wonder and Muse Wonder Full, the company's first self-described frontier models. The release caps roughly fourteen months of AI overhaul that began after Mark Zuckerberg put Scale AI co-founder Alexandr Wang in charge of the effort. Both models were trained at Prometheus, Meta's gigawatt-scale data center in central Ohio that came online earlier this year, and they arrive barely six weeks after the previous internal best, Muse Spark 1.3. Reporters at TechCrunch, The Verge and WIRED noted this is Meta's first credible answer to the state-of-the-art releases that OpenAI and Anthropic have treated as routine, and its first flagship push since Llama 4 drew a rocky reception in 2025.
Meta's announcement calls the pair "our most capable models to date," built for "personal and professional agentic tasks." The model card says Wonder and Wonder Full post benchmark scores comparable to Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra, and that on the MemoryBench recall test the standard model outscores Fable, Opus and Astra alike. Meta is explicit that these are its own numbers: a full safety report with additional evaluations is promised for later. CNBC reported that the models carry the internal codename "Watermelon."
Artificial Analysis, an independent benchmarking firm, is the harder yardstick. Its Intelligence Index v4.3, a composite of ten evaluations, puts Claude Opus 5.5 at 58, Muse Wonder Full at 56, Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon at 53, and standard Muse Wonder at 52 — a tie with GPT-6.1 Sol and an 8-point jump from Muse Spark 1.3's 48. On individual tests the picture is mixed: the standard model moved from 33% to 52% on Terminal-Bench 4.0 and from 58% to 68% on AutomationBench-AA, while Meta's own charts show Wonder Full at 57.5% on SWE-Marathon against Claude Opus 5.5's 46.3%. Opus 5.5 still leads ProgramBench 70.5 to 59.0 and Terminal-Bench 4.0 59.6 to 57.1.
Rather than chase raw scores alone, Meta is selling two mechanisms aimed at agents: Active Memory, which carries relevant context about a user across conversations for days, weeks or months, and Active Focus, which pushes the model to try alternative routes when it hits a blocker instead of stalling. The company's examples are multi-step work — booking reservations, interpreting leases, building applications — plus professional domains including law, finance and health. Wang told CNBC that training deliberately rewarded "trying other approaches instead of stalling and then checking back in with the user at any time."
The Meta Model API lists muse-wonder-1.0 at $4 per million input tokens and $20 per million output tokens, roughly the sticker price of GPT-6 Sol and Opus 5.5. Artificial Analysis calculates that each Intelligence Index task costs $4.52 on Wonder, 2.8 times Muse Spark 1.3's $1.60 but still below Claude Fable 5.1 ($7.63), Claude Opus 5.5 ($5.98) and Claude Sonnet 5.5 ($5.46), and above GPT-6 Astra ($3.26), Gemini 4 Argon ($1.99) and GPT-6.1 Sol ($0.72). Meta's chief AI officer said API access will be priced "meaningfully below" leading models from OpenAI and Anthropic.
Availability is narrow. Wonder went live Monday through the Meta Model API and Muse Code, with 3,000 requests per minute and 4 million tokens per minute rate limits reported for API users. The Muse agent app, which launched September 8, will get the model later, and only for subscribers paying $20 or $100 a month. Wonder Full is restricted to Muse Code's $50 "power usage" and $200 "maximum usage" tiers and has no published API pricing. Open weights are not part of the release.
Zuckerberg said on X last month that the Muse models had been held back to work on safety issues, and Meta published its Superintelligence Scaling Framework earlier in October. The New York Times reported days before launch that Zuckerberg argued in an August meeting for shipping Muse despite known test concerns, with Wang and AI product head Nat Friedman aware of them; a Meta spokesman disputed that external pressure drove the timing, saying the company "even delayed shipping Muse for several months to make sure we got this right." Earlier safety wobbles included Muse handing a user's address to a Facebook Marketplace seller without permission.
Three things will decide whether this is a genuine frontier entry: when the full safety evaluation lands, when Wonder reaches paying Muse subscribers, and whether Meta opens the weights. TheNextWeb observed that the open-weight question determines how Meta is classified in Europe, and the company has not said. Meta has also promised specialized models built on this work — the clearest signal yet that the fourteen-month rebuild is meant to produce a steady flagship cadence, not a one-off release.