Meta Releases Muse Wonder Models, Claiming Frontier AI Status as Early Users Report Mixed Results

The models, developed under the internal codename Watermelon, target personal and professional agent work, with Meta emphasizing cross-conversation memory, long-duration task execution and coding. Chief executive Mark Zuckerberg said in an interview that Meta has reached the frontier of AI technology. According to Meta's published evaluations, Muse Wonder Full is competitive with GPT-6.1 Astra and Claude Fable 5.1 on at least some benchmarks. Third-party testing by Artificial Analysis found the Wonder series a marked improvement over the previous generation, placing in the top three in several agent and professional-work evaluations, according to the technology outlet ifanr.

The launch follows more than a year of internal rebuilding. After Llama 4 fell short in April 2025, Zuckerberg increased spending on talent and computing, recruited Scale AI founder Alexandr Wang and reorganized the research team. In April this year the reorganized team shipped Muse Spark, which still trailed rivals. Meta says the team spent the past 14 months rebuilding its technical stack from the bottom up and scaling pretraining and reinforcement learning; the Wonder models were trained in the gigawatt-scale Prometheus data center.

Two features carry the release. Active Memory keeps important information across conversations and retrieves it when a new task needs it, so preferences, constraints and personal details raised earlier can inform later work instead of being restated each time. Wang told Axios that personal agents need important information to persist across days, weeks and months of exchanges. Active Focus addresses tasks that stall halfway: pages change, information is missing, bookings fail, and the model is expected to try other routes, adjust its plan and keep going. Meta says the model will still ask users when necessary and seek confirmation before important actions. Zuckerberg described the approach as “the model is the product,” saying capabilities and personality belong in training. Wang added that when an agent should act on its own and how it speaks to users are product behaviors requiring dedicated design, and gave CNBC the example of booking a restaurant, where the agent may hit obstacles while hunting for a table and should try other paths rather than immediately asking for help.

Coding serves as the test case. A continuously developed software project requires the model to hold requirements in mind, read existing code, change multiple files and iterate through testing. Meta's demonstrations included turning a photo into a playable 3D kitchen game and assembling a running plan that draws on personal preferences while coordinating a browser and several apps. For professional work, Meta listed legal, financial and health evaluations, among them Harvey's Legal Agent Benchmark, which requires models to produce legal work products from source materials and detailed instructions.

Artificial Analysis found the Full version placed in the top three in six of ten tests. At maximum reasoning effort, Muse Code paired with Wonder Full matched Claude Code with Sonnet 5.5 and edged out the pairing with Opus 5.5.

Early reactions on social platforms were mixed. User Chetaslua posted an animation featuring Dragon Ball characters, said the visuals were generated entirely from code and praised its response speed, noting the voiceover came from ElevenLabs and the music from Suno. User Max Blade gave Muse Wonder Full and Opus 5.5 the same prompt for an “AAA lawnmower simulator” played while drinking beer and smoking cigars, and said Opus rendered grass, trees and lighting in detail while Muse's scene and hand modeling were coarse. User Bhavy, comparing rocket scenes, said Muse Wonder's output was worse than Kimi K3 while his recorded cost reached $14.30 against about $6.90 for Kimi.

Wang acknowledged on social media that the new model's coding ability still trails Opus 5.5 and that the team is investigating several bugs. He said the team did not specifically train for such demos and wanted users to apply the model to real programming work, adding that training directions were chosen around actual Muse user needs while conceding the demos could be improved.

Meta's official blog also addressed whether gains transfer to tasks not emphasized in training. The team compared Muse Spark 1.3 and Muse Wonder Full on familiar and unfamiliar tasks, saying both improved as reasoning compute increased but that Wonder Full performed better on unfamiliar tasks than the Spark trend would predict, which Meta attributed to a larger and stronger base model.

Muse Wonder is available through the Meta Model API and Muse Code, while the more capable Wonder Full is limited to Muse Code subscribers. CNBC reported that Muse Code tiers cost $50 and $200 a month. The Muse personal assistant app will receive Wonder later, with access prioritized for subscribers. Muse offers no model selector and no reasoning-effort options; Wang argues that ordinary users should not have to study technical manuals. API pricing is $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. Both models support a 1 million token context window and accept text, images, video and audio, returning text.

On the Artificial Analysis Intelligence Index, Muse Wonder at maximum reasoning effort cost about $4.52 per task, below Claude Fable 5.1, Opus 5.5 and Sonnet 5.5 in the same comparison but above GPT-6 Astra at $3.26, Gemini 4 Argon at $1.99 and GPT-6.1 Sol at $0.72. Artificial Analysis said Wonder uses fewer output tokens than the previous generation, yet higher input consumption and API pricing push up the total. Blade said Muse's appetite exhausted the quota in his $50 plan, forcing him to upgrade to the $200 tier to finish the task.

On open source, Wang said Meta's commitment is unchanged, citing the previously released Muse Glimmer weights and technology for deploying Muse agents on hardware. CNBC noted Meta is trying to open a revenue stream beyond its core advertising business.

Some users have complained on social platforms that Muse accessed personal messages without authorization. Meta disputes this, saying some cases may involve users granting permission in earlier conversations and forgetting, and that the team is tracking the issues and helping users understand how authorization takes effect. Meta said the Wonder series strengthened defenses against malicious instructions, the handling of private information and the prevention of cyber and biological misuse, with a fuller safety report to follow. Zuckerberg mentioned an unreleased Muse Confidential VM intended to keep user data and agent conversations opaque even to Meta. He told the Financial Times that he rejects putting the brakes on AI development and that labs should ensure safety through technical measures and responsible development, adding that Meta is already working on next-stage models and expects progress to accelerate.