Meta releases the Muse Wonder series of models, focused on personal agent memory and long-horizon tasks

[[**Meta releases the Muse Wonder series of models (internal codename "Watermelon"), focused on personal agent tasks; the company claims some benchmarks are on par with GPT-6 Astra and Claude Fable 5.1, but the AI lead admits coding ability lags behind Opus 5.5, and early user reception is clearly mixed.**

**Key points**

**1. Two core upgrades: memory and execution**

Active Memory retains user preferences and constraints across conversations, while Active Focus lets the model try alternative approaches when a task hits a snag instead of immediately asking for help. This aims to solve the two big pain points of personal assistants: "can't remember" and "can't finish."

**2. Official results are impressive, third-party evaluation ranks in the top three**

Meta's evaluation shows Wonder Full matches top competitors on multiple benchmarks, and Artificial Analysis testing also shows it broke into the top three on six of its ten tests. The performance gains come from training completed at the gigawatt-scale Prometheus data center.

**3. Coding ability is recognized, but the 3D demo flops**

Muse Code paired with Wonder Full's coding score is on par with Claude Code paired with Sonnet 5.5 and slightly beats Opus 5.5. However, in early hands-on experience the 3D game task results were crude, and the company admits its coding ability still lags behind Opus 5.5.

**4. Pricing is lower than top competitors, with an obvious cost advantage**

API pricing is $4 per million input tokens and $20 per million output tokens; at the highest reasoning effort a task costs about $4.52, lower than Opus 5.5 and Sonnet 5.5 but higher than GPT-6 Astra. Wonder Full is available only to subscribers, and users need to upgrade to the $200 plan to complete complex tasks.

**5. Privacy controversy and the debate over the open-source path**

Some users complained that Muse accessed personal messages without authorization; Meta responded that this may be because users authorized it earlier and forgot. Although the Wonder series is closed-source, Meta says it still supports the open-source path, seeking new revenue beyond advertising in subscriptions, APIs and the enterprise market.**Meta releases the Muse Wonder series of models, focused on personal agent memory and long-horizon task execution, with some benchmarks matching top models such as GPT-6 Astra; but early reception is mixed, coding and 3D capabilities are questioned, and controversy over commercial pricing and privacy authorization has also surfaced.**

**Key points**

**1. The new models focus on memory and continuous execution**

The Wonder series introduces Active Memory to retain important information across conversations, and Active Focus to try alternative paths when a task is blocked, aiming to improve the continuity and proactivity of personal assistants handling real-world matters.

**2. Benchmark performance ranks among the leaders**

In Artificial Analysis evaluations, Wonder Full entered the top three on six of ten tests; at the highest reasoning effort, Muse Code paired with Wonder Full's coding agent score is on par with Claude Code paired with Sonnet 5.5 and slightly beats the combination paired with Opus 5.5.

**3. User testing results and reception are mixed**

Some users praised the fast code generation responses, but 3D game scenarios were clearly crude; Meta's AI lead admitted the new model's coding ability is inferior to Opus 5.5, explaining that it was not specifically trained for demo scenarios and hoping users would focus on real coding tasks.

Wonder Full is available only to Muse Code subscribers; API pricing is $4 per million input tokens and $20 per million output tokens; according to the Intelligence Index, the cost per task is about $4.52, higher than competitors such as GPT-6 Astra. Some users, having exhausted their $50 plan quota, were forced to upgrade to the $200 plan.

**5. The privacy authorization controversy has not subsided**

Users complained that Muse accessed personal messages without authorization, while Meta said that in some cases this may stem from authorization already granted in earlier conversations, and the team is tracking the issue; a more complete security report will be released later, and there are plans to launch a Confidential VM solution that is kept secret even from Meta itself.2026-10-13 07:04 Just now, Zuckerberg released the mysterious "Watermelon" model, benchmark scores rival Opus 5.5, but reception is hard to describe in one word AppSo Quick Look

This article is from the WeChat public account: APPSO, author: the one who discovers tomorrow's products, original title: "Just now, Zuckerberg released the mysterious 'Watermelon' model, benchmark scores rival Opus 5.5, but reception is hard to describe in one word"

Last year, Zuckerberg waved his checkbook around the world recruiting AI talent, with some compensation packages reaching hundreds of millions of dollars, a move as grand as an NBA star transfer. Now, the expensively assembled team has finally delivered a report card closer to expectations.

Just now, Meta released the new Muse Wonder series models (internal codename Watermelon), including Muse Wonder and Muse Wonder Full, focused on personal and professional agent tasks, with emphasis on improving cross-conversation memory, long-duration task execution, and coding capabilities.

In an interview, Zuckerberg said that Meta has reached the forefront of AI technology.

According to the benchmarks Meta published, the new Muse Wonder Full model can at least hold its own against GPT-6 Astra and Claude Fable 5.1 on some benchmarks.

Tests by the third-party organization Artificial Analysis also show that Wonder Full is a complete transformation from the previous generation, entering the top three in multiple agent and professional work evaluations.

The team assembled at great expense has given Meta the confidence to challenge the leading models again, and has also made Meta Muse's path as a Personal Agent smoother.

This Meta release continues more than a year of restructuring its AI business.

In April 2025, Llama 4 stumbled, and Zuckerberg immediately increased investment in talent and compute, recruited Scale AI founder Alexandr Wang, and reorganized the research team. This April, the reorganized team launched Muse Spark, but it still lagged a step behind competitors.

According to Meta, over the past 14 months the team tore everything down and rebuilt the technology stack from the ground up, while continuously scaling up pretraining and reinforcement learning. The Wonder series was trained in the gigawatt-scale Prometheus data center, making it the latest chip cashed in from this round of high-stakes betting.

Muse Wonder makes changes around two directions: Active Memory and Active Focus.

Active Memory is responsible for retaining important information across conversations and proactively retrieving relevant content when a new task requires it. Preferences, constraints, and personal circumstances the user mentioned earlier can serve as references for subsequent tasks, reducing the hassle of restating background every conversation.

Wang told Axios that a personal agent needs to let important information persist across days, weeks, or even months of interaction.

Swipe left to see more content

Active Focus targets the problem of stopping halfway through task execution. Real-world operations encounter page changes, missing information, or failed bookings, and the model needs to try other methods, adjust plans based on new information, and keep pushing the task forward. Meta says the model will still ask the user when necessary and seek confirmation before important operations.

In a recent interview, Zuckerberg summarized this approach as "the model is the product," with product features such as capabilities and personality all needing to go into the training process. Wang added that when an agent should act proactively and how it communicates with users are also product behaviors that require dedicated design.

Wang gave CNBC the example of booking a restaurant. An agent querying availability may encounter various obstacles, and one training goal is to make it try other paths, reducing the situation where it stops and asks the user for help as soon as it hits a problem.

Coding is the touchstone for testing these capabilities.

A software project under continuous development requires the model to keep track: remember requirements, understand existing code, modify multiple files, and repeatedly test and fix. Meta says both models have been trained on various long-duration tasks to improve their ability to keep going.

The official demonstration also includes "cooking from a picture," generating a playable 3D kitchen game from an image, and arranging a running plan by combining personal preferences and mobilizing the browser and multiple apps.

For professional work, Meta listed law, finance, and health-related evaluations, among which Harvey's Legal Agent Benchmark requires the model to generate legal work products based on source materials and detailed instructions.

Artificial Analysis's evaluation shows that the Wonder series' overall capability has improved markedly, with the Full version entering the top three in six of ten tests. At the highest reasoning intensity, the coding agent score of Muse Code with Wonder Full is on par with Claude Code with Sonnet 5.5, and slightly ahead of Opus 5.5.

Early hands-on impressions on social platforms are a mix of joy and frustration.

User Chetaslua showed an animation featuring Dragon Ball characters, saying the visuals were generated entirely by code, and gave a thumbs-up to Muse's response speed. He also noted that the voiceover came from ElevenLabs and the music was made by Suno.

Another user, Max Blade, gave Muse Wonder Full and Opus 5.5 the same prompt: make a "AAA-grade lawnmower simulator" where you can drink beer and smoke a cigar.

From the comparison footage he posted, the Opus version's grass, trees, and lighting are rendered in fine detail, while the Muse version's scene and character hands are visibly rough. He believes Muse still falls short on this kind of 3D game task.

A rocket scene comparison shared by user Bhavy drew similar feedback. He said the Muse Wonder result was not as good as Kimi K3, yet the cost he recorded reached $14.3, while Kimi was about $6.9, amounting to much effort for half the result.

Meta AI head Alexandr Wang also acknowledged when responding to users on social media that the new model's coding ability still lags behind Opus 5.5, and that the team is investigating multiple bugs.

Regarding the 3D game demo results that some users were unhappy with, Wang explained that the team did not specifically train for this kind of demo, and he hopes users will apply the model to real, serious coding work. He said the team chose its training direction based on Muse users' actual needs, while acknowledging that it could put more effort into demo scenarios.

In its official blog, Meta also tried to answer another question: whether the model's progress can generalize and transfer to tasks that were not a focus of training.

The team compared Muse Spark 1.3 and Muse Wonder Full on "familiar" and "unfamiliar" tasks. According to Meta's explanation, as inference compute increases, both models improve, but Wonder Full's performance on unfamiliar tasks still exceeds what would be expected by extrapolating from Spark's trend by a notable margin.

Meta attributes this to a stronger foundation: a larger, more capable base model allows the benefits of post-training to carry over to related tasks.

For personal agents that need to handle different goals, tools, and constraints, this transferability is especially important. However, these conclusions come from Meta's own "homework," and exactly which real-world scenarios they cover will need more time to be verified.

Beyond the leaderboards, Meta also has to convince users to keep paying long-term

Alongside the model capability upgrade, Meta is also striking while the iron is hot to push paid offerings.

At launch, Muse Wonder was already available through the Meta Model API and Muse Code, while the more capable Wonder Full is offered only to Muse Code subscribers.

The Muse personal assistant app used by ordinary users will later integrate Wonder and will be opened first to subscribers.

On the API side, Muse Wonder is priced at $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, can accept text, image, video, and audio input, and output text.

According to Artificial Analysis's Intelligence Index test, at the highest reasoning intensity, Muse Wonder costs about $4.52 per task, lower than Claude Fable 5.1, Opus 5.5, and Sonnet 5.5 in the same comparison, but higher than GPT-6 Astra at $3.26, Gemini 4 Argon at $1.99, and GPT-6.1 Sol at $0.72.

Artificial Analysis noted that although Wonder saves on output tokens compared with the previous generation, higher input consumption and API pricing still push up the total bill. What users ultimately have to calculate also includes how many tokens it takes to complete the same job, and how many times they have to redo the work along the way.

In the aforementioned lawnmower simulator test, Max Blade labeled the "API-equivalent cost" for Muse Wonder Full and Opus 5.5 as $4.11 and $99.43, respectively, but he said Muse has such a big appetite that it exhausted the available quota of his $50 plan, forcing him to upgrade to the $200 plan to finish the task.

Of course, as the big brother of open-source models, this launch has also put Meta's "open-source persona" under scrutiny again. Although Muse Wonder adopts a closed-source model, Wang said Meta's original aspiration remains unchanged and it still supports open source, noting that it had previously opened the weights of Muse Glimmer and related technologies for deploying Muse agents on hardware.

Compared with its previous approach of trading openness for a circle of friends, Wonder has Meta entering the field more directly, fighting at close quarters in subscriptions, APIs, and the enterprise customer market. CNBC noted that Meta is trying to open another revenue stream beyond its core advertising business.

But whether agents can grow into a money tree still depends on variables beyond the model.

Recently, some users complained on social media that Muse accessed personal messages without authorization. Meta disputed this, saying that in some cases users may have granted authorization in earlier conversations and then forgot about it; the team is tracking these issues and helping users understand how authorization took effect.

The two sides each stick to their own account, but the product design of Personal Agent must face a reality: users may have nodded in agreement long ago, yet the model will use that old consent as a license in later tasks, so the scope and validity period of authorization therefore need to be clear at a glance.

Meta said the Wonder series has strengthened its ability to resist malicious instructions, handle private information, and prevent abuse in the cyber and biological fields, and that a more complete safety report will be released later.

Zuckerberg mentioned the not-yet-launched Muse Confidential VM, saying the plan is intended to make user data and conversations with agents opaque even to Meta itself.

Interestingly, in an interview with the Financial Times, Zuckerberg rejected the idea of "hitting the brakes" on AI development, arguing that labs should ensure safety through technical measures and responsible development. He also revealed that Meta is already developing the next stage of models and expects progress to be even faster afterward.

Continuously improving model capabilities must ultimately support the Personal Agent envisioned by Zuckerberg, enabling AI to handle personal daily affairs. Using the most cutting-edge technology to order food, arrange itineraries, and handle emails sounds like overkill, but these tasks all involve constantly changing information and differing personal requirements.

The direction Muse Wonder has chosen also gives model progress a more concrete way to be measured. AI capability can be tested by the hardest problems, while the value of a personal agent must be proven by those small, everyday matters that repeat every day and always need someone to worry about them. AI industry signal channel: Frontier technology. This content comes from the internet, original link; the views represent only the author and do not represent Huxiu's position. For copyright issues, please contact hezuo@huxiu.com, and we will verify and handle it promptly. Everyone who is changing the world and wants to change the world is on the Huxiu APP[ ]