Meta Muse Wonder: First Frontier Model for Enterprise Agents
Meta Muse Wonder Launches: A Frontier Model Built for Enterprise Agents
October 12, 2026. Meta Superintelligence Labs today released Muse Wonder and Muse Wonder Full, the company’s first frontier-class AI models. Available immediately through the Meta Model API and Muse Code, the Wonder series represents what Meta calls “a major leap toward personal superintelligence,” with benchmark results it says match or surpass the best models from Anthropic and OpenAI.
For enterprise teams, the more immediate story is the two new capabilities arriving at the frontier tier: Active Memory and Active Focus, both designed to make AI agents useful across the multi-step, multi-day workflows that define real enterprise work.
What Happened
Meta CEO Mark Zuckerberg restructured the company’s AI efforts fourteen months ago, forming Meta Superintelligence Labs and hiring former Scale AI CEO Alexandr Wang to lead it. The effort produced a succession of models: Muse Spark (April 2026), then 1.1, 1.2, and 1.3, each compressing the gap to the frontier. Muse Wonder, previously known internally as Watermelon, is the result.
Where the Muse Spark series was a mid-tier frontier model, Wonder enters a different category. TechCrunch describes it as “Meta’s first attempt at a frontier model,” and Axios reports Alexandr Wang calling them “true frontier models, competitive with the best models in the world.”
Both models were trained at Meta’s Prometheus data center in central Ohio, a gigawatt-scale facility that came online earlier this year. Zuckerberg told the Financial Times that “getting to this point both required a large amount of compute and investment, and a lot of very rigorous scientific work” from the superintelligence team.
Active Memory and Active Focus: Why They Matter for Enterprise
Most enterprise agent failures trace to one of two causes: the agent forgets prior context, forcing users to re-explain their situation at the start of every conversation, or the agent abandons complex tasks when conditions become uncertain, returning a partial answer instead of pushing through blockers.
Active Memory addresses the first. According to Meta’s research blog, this capability carries what matters to a user “across conversations over days, weeks, months and beyond,” and actively retrieves the relevant details when a new task calls for them. For enterprise agents serving the same users repeatedly, such as an internal legal assistant or a procurement co-pilot, persistent memory is a prerequisite for genuine usefulness rather than a convenience feature.
Meta evaluated Active Memory using MemoryBench, a continual learning benchmark. Wired reports that Muse Wonder now scores higher on MemoryBench than Claude Fable 5.1, Claude Opus 5.5, and GPT-6 Astra, making it the leading model on this specific capability as of launch.
Active Focus targets the second failure mode. Meta’s company blog post describes it as the model’s ability to “keep working toward your goal when a task gets hard,” trying alternative approaches, recovering from errors, and adapting to new information rather than escalating prematurely or stopping silently. The model checks with users before taking consequential actions, reducing the risk of compounded errors on long-running tasks.
Meta tested Active Focus through FocusBench and SWE-Marathon, which evaluates sustained work on long-running software projects. Together, Active Memory and Active Focus change the practical architecture of enterprise agents built on Muse Wonder compared to those built on earlier, stateless models.
Benchmark Position: Frontier Tier, Not Below It
Meta published direct comparisons against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across domain expertise (law, finance, health), coding, and memory benchmarks.
Wired’s analysis summarizes: Wonder Full “ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index,” while outperforming GPT-6 Astra and Claude Fable 5.1 “on some specialized benchmarks, including ones that measure a model’s ability to work with legal documents and to build software.”
| Benchmark | Muse Wonder Full position | Source |
|---|---|---|
| Artificial Analysis Intelligence Index | Just below Opus 5.5 | Wired, Oct 12, 2026 |
| MemoryBench | First (above Fable 5.1, Opus 5.5, Astra) | Meta / Wired confirmed |
| Harvey’s Legal Agent Benchmark | First among tested models | Meta published |
| SWE-Marathon (coding, sustained) | Competitive with frontier leaders | Meta published |
The Muse Spark 1.3 generation peaked at Intelligence Index score 62. Muse Wonder Full clears that ceiling and competes directly with the models Anthropic and OpenAI use to anchor their enterprise offerings.
For enterprise teams currently choosing between GPT-6 Astra and Claude Opus 5.5, a third credible option at the same tier now exists, with a distinct profile: stronger on memory and legal tasks, slightly behind on overall general intelligence.
Pricing and Availability
Muse Wonder is available immediately through the Meta Model API, priced at $4.00 per million input tokens and $20.00 per million output tokens. TechCrunch notes this places it “roughly equivalent in cost to GPT-6 Sol or Opus 5.5,” firmly in the frontier pricing tier rather than the mid-tier pricing of Muse Spark.
Muse Wonder Full, the more capable variant, is available only to Muse Code subscribers at $50/month (power) or $200/month (maximum). It is not accessible through the API. Meta plans to use Wonder to power its Muse personal agent, rolling out first to paid subscribers, though no specific timeline was given.
On open weights: Wang told Axios that Meta “remains committed to open source” and has previously released Muse Glimmer. No commitment or timeline exists for Wonder-family weights.
Safety Considerations for Enterprise Deployment
Meta’s announcement emphasizes that Muse Wonder includes “our strongest safeguards to date against CBRN and cyber misuse,” with specific attention to the risks introduced when an agent accesses personal messages, calendars, and private documents on a user’s behalf.
Wired flags a real tension: “Agents that remember personal information and act on a user’s behalf also introduce new privacy and security challenges.” Meta did not disclose specific red-teaming results for Wonder-powered agent behavior, though it noted the models were trained against a range of misuse categories and that a fuller safety report is forthcoming.
For enterprise security teams, persistent-memory models require a different governance posture than stateless models. The multi-agent governance frameworks that organizations are deploying now will need to account for how long-lived agent memory stores are audited, segmented, and cleared.
What Enterprise Teams Should Do Now
The Muse Wonder launch changes the frontier model map in two concrete ways.
First, enterprise teams evaluating models for agentic workflows now have a fourth serious vendor at the top tier. Anthropic Opus 5.5, GPT-6 Astra, and Google Gemini 4 Argon now share the frontier table with Meta. Wonder’s distinct strengths in memory and legal domain benchmarks make it a priority evaluation candidate for legal, procurement, compliance, and finance use cases.
Second, Active Memory represents a capability that most enterprise deployments have been working around using external vector stores and session-level context injection. A model with trained, built-in persistent memory changes the architecture options for teams building long-running agent workflows and can reduce the engineering overhead required to make agents feel coherent across sessions.
The practical first step: run your current production workloads through Muse Wonder on the API, compare against your incumbent model on the tasks that matter for your operation, and use MemoryBench-style evaluations to verify whether Active Memory actually holds up in your domain. Pricing is comparable to Opus 5.5 and GPT-6 Astra, so the evaluation does not require a pricing trade-off.
If you are structuring an enterprise agentic strategy that depends on agents carrying context across sessions, the planning work starts here.