Meta Debuts Muse Wonder AI Models in Proprietary Shift
Developed under Alexandr Wang's Meta Superintelligence Labs, the proprietary Muse Wonder models match frontier rivals on coding and agent workflows.
In short
- Meta launched two closed-source frontier AI models on October 12, 2026: Muse Wonder and Muse Wonder Full.
- Muse Wonder Full reached 56 on the Artificial Analysis Intelligence Index, tying Claude Sonnet 5.5 and trailing only Claude Opus 5.5.
- Meta broke from its open-weights tradition, offering standard Muse Wonder via commercial API at $4/$20 per million tokens and locking Muse Wonder Full behind Muse Code subscriptions.
- The models introduce Active Memory and Active Focus to handle long-running agentic tasks across software engineering, legal review, and business workflows.
Meta entered the top tier of frontier AI on Monday, October 12, 2026, officially releasing its proprietary Muse Wonder and Muse Wonder Full models. Built by Meta Superintelligence Labs, the new releases target complex reasoning, professional agent workflows, and software development. Following earlier coverage of the system, neither Muse Wonder model includes publicly downloadable weights, marking a strategic pivot away from the company's historical open-weight Llama releases toward commercialization and closed infrastructure.
The launch marks the culmination of a 14-month overhaul of Meta's artificial intelligence group. Under Meta Superintelligence Labs (MSL), established after Meta acquired Scale AI for $14.3 billion and named Alexandr Wang chief AI officer, the company spent tens of billions of dollars on compute and engineering talent. According to Meta's official announcement, both models were trained using large-scale pretraining and reinforcement learning at Meta's gigawatt-scale Prometheus data center.
Core Architecture and Agentic Features
Both Muse Wonder models accept inputs across text, images, audio, and video, returning text responses with a context window of approximately 1,000,000 tokens. Mark Zuckerberg described Wonder on X as a much larger and more advanced system than the earlier Muse Spark family released earlier in the year. To address failure points common in automated agent execution, Meta designed the models around two primary software mechanisms: Active Memory and Active Focus, as reported by CNET.
Active Memory allows the agent to retain personal and professional context across multiple conversations, recalling previous interactions and applying prior constraints to active tasks. Active Focus guides long-horizon execution. According to Meta's research blog, this system allows the model to try alternative approaches when blocked, recover from software execution errors, revise plans, and ask users for confirmation before completing irreversible actions.
Meta demonstrated these tools running tasks such as generating working 3D software from images and building a running plan across web applications, while highlighting use cases such as parsing apartment leases and handling insurance claims.
Benchmark Performance Against Frontier Competitors
Independent evaluations and Meta's model card place Muse Wonder directly alongside flagship offerings from OpenAI, Anthropic, and Google. On the aggregate Artificial Analysis Intelligence Index, which combines ten evaluations across reasoning, coding, and knowledge work, Muse Wonder Full scored 56, according to evaluation data published by VentureBeat. That ties Anthropic's Claude Sonnet 5.5 (56) and sits behind Claude Opus 5.5 (58), while topping OpenAI's GPT-6 Astra (53), Google's Gemini 4 Argon (53), Anthropic's Claude Fable 5.1 (53), and standard Muse Wonder (52), as also noted by Mashable.
In domain-specific benchmarks reported in Meta's model card, Muse Wonder Full demonstrated clear strengths in coding and legal evaluation while lagging behind competitors in system terminal control and medical knowledge:
- Software Engineering (DeepSWE 1.1): Wonder Full achieved 79.0 and standard Wonder reached 78.8, outperforming GPT-6 Astra (74.1) and Claude Opus 5.5 (74.2).
- Long-Running Coding (SWE-Marathon): Wonder Full scored 57.5 and standard Wonder scored 53.1, ahead of Claude Opus 5.5 (46.3) and GPT-6 Astra (42.5).
- Legal Tasks (Harvey LAB-AA): Wonder Full reached 13.1 and standard Wonder scored 10.0, topping GPT-6 Astra (8.6) and Claude Opus 5.5 (4.2), though Meta noted that low absolute scores across all models reflect the difficulty of legal deliverable generation.
- Terminal Environment Tasks (Terminal-Bench 4.0-AA): Claude Opus 5.5 led with 59.6, followed by GPT-6 Astra at 59.1, Wonder Full at 57.1, and standard Wonder at 52.0.
- Clinical Knowledge (HealthBench Pro): Claude Opus 5.5 posted 65.6 and GPT-6 Astra reached 65.3, while Wonder Full scored 62.3 and standard Wonder scored 60.5.
Meta noted that some software engineering evaluations involved different tool configurations rather than strictly identical environments, and third-party benchmark evaluations were combined with internal evaluations.
Pricing, API Availability, and Closed Distribution
The distribution structure of the Wonder family marks an operational departure for Meta. As reported by VentureBeat, neither model has been released under an open-source license, unlike Meta's smaller Muse Glimmer model that uses Apache 2.0. Instead, access is split between a paid developer API and proprietary subscriptions.
Standard Muse Wonder is available through the Meta Model API under the identifier muse-wonder-1.0. Rates are set at $4.00 per 1 million input tokens, $20.00 per 1 million output tokens, and $0.20 per 1 million cached input tokens, matching the standard pricing of Claude Opus 5.5. Web search grounding is priced at $2.50 per 1,000 queries, with rate limits capped at 3,000 requests per minute and 4 million tokens per minute. Artificial Analysis calculated an estimated average task cost of $4.52 for Muse Wonder, compared to $3.26 for GPT-6 Astra and $5.98 for Claude Opus 5.5.
By contrast, the higher-performing Muse Wonder Full is completely excluded from pay-as-you-go API access. It is offered exclusively within Meta's Muse Code programming environment, as confirmed by Engadget. Developers must subscribe to either the Power Usage tier at $50 per month or the Maximum Usage tier at $200 per month. Users cannot export Wonder Full weights or integrate the model directly into standalone enterprise applications.
Data Governance and Safety Considerations
Meta established specific data retention tiers for the new models. Pay-as-you-go API calls operate under Standard Mode, meaning Meta does not use prompts or completions to train future models. Approved enterprise API customers can also access Zero Data Retention (ZDR), which disables persistent storage but turns off server-side conversation state, web search, background execution, and file uploads.
Conversely, Muse Code subscriptions default to Contributor Mode, granting Meta permission to use submitted code, files, contextual data, and model outputs for future training. While users can turn off Contributor Mode, the opt-out applies only to subsequent submissions, and Meta's terms explicitly prohibit submitting confidential or personal information while the setting remains active.
According to Meta AI's post, Meta included safety evaluations for chemical, biological, radiological, and nuclear (CBRN) hazards as well as autonomous cyber operations, promising a comprehensive alignment report soon. The company stated that human verification remains recommended before deploying Wonder agents for high-consequence business actions.
Questions people ask
Are the weights for Muse Wonder or Muse Wonder Full available for download?
No. Both Muse Wonder and Muse Wonder Full are closed, proprietary models. Neither model has downloadable parameters for self-hosting, departing from Meta's traditional open-weight Llama releases.
How can developers access Muse Wonder and Muse Wonder Full?
Standard Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is available only to paid subscribers of Muse Code on plans priced at $50 or $200 per month.
What are Active Memory and Active Focus?
Active Memory is Meta's feature for retaining context and relevant user details across separate chat conversations. Active Focus enables the model to persist through software execution roadblocks, recover from errors, test alternatives, and ask for user clarification during complex tasks.
Sources
- Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via APIventurebeat.com, Oct 12, 2026
- Meta Says It's Caught Up To The Frontier With Its Latest Wonder AI ModelsEngadget, Oct 12, 2026
- Muse Wonder: A Frontier Model For Personal & Professional AI AgentsMeta AI, Oct 12, 2026
- Meta Launches Muse Wonder AI Models to Play Catch-Up With Rivals - CNETCNET, Oct 12, 2026
- Meta launches Muse Wonder, catching OpenAI, Anthropic: How to try it, benchmark resultsMashable, Oct 12, 2026
How this story was made: the newsroom picked it up from Google News and Reddit, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (37 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 13, 2026 at 02:45 UTC
