Analyst(s): Nick Patience
Publication Date: October 14, 2026
Meta has entered the frontier model market with Muse Wonder and Muse Wonder Full, its first models to rank alongside OpenAI, Anthropic, and Google on independent benchmarks. Wonder leads on memory and long-running agent tasks, although closed weights, a subscription-only flagship, and personal data risk will shape how far it reaches enterprise buyers.
What Is Covered in This Article:
- Meta launched Muse Wonder and Muse Wonder Full, its first frontier models since rebuilding its AI organization, with Wonder priced at $4/$20 per million input/output tokens on the Meta Model API.
- Independent benchmarks place Wonder level with OpenAI’s GPT-6 Astra and behind Anthropic’s Claude Opus 5.5 and Sonnet 5.5, with its clearest strengths in memory, long-running coding, and a low hallucination rate.
- Both Wonder models are proprietary and closed-weight, and the stronger Wonder Full is available only through Muse Code subscriptions, a break from Meta’s Llama-era approach.
- Against Google’s Gemini 4 Argon, announced in September, Wonder is level on general intelligence but trails on legal work, hallucination, and enterprise distribution.
- The frontier model market is expanding and converging, with closely matched capabilities and standard pricing settling at $4/$20.
The News: Meta Superintelligence Labs has released Muse Wonder and Muse Wonder Full, the first models that Meta positions at the frontier of AI capability. Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with a 1 million-token context window and cached input at $0.20 per million tokens. The more capable Muse Wonder Full is not offered through the API and is available only through Muse Code subscriptions at $50 or $200 per month. Meta says both models will power its Muse personal agent app, starting with subscribers on $20 and $100 monthly plans. Meta highlights two capabilities it calls Active Memory, which carries context about a user across conversations, and Active Focus, which keeps an agent working through long, multi-step tasks.
Meta trained both models at its gigawatt-scale Prometheus data center in Ohio and describes them as the product of a 14-month overhaul of its AI organization. It published an eight-page model card at launch and says a fuller safety report will follow. The release follows Google’s announcement of Gemini 4 Argon and OpenAI’s launch of GPT-6.1 Sol, both in September 2026.
Meta Muse Wonder Reaches the Frontier. Can Meta Sell It?
Analyst Take: The frontier model market spent most of 2026 in a contest between OpenAI and Anthropic, and the gap has widened quickly. Meta Muse Wonder is a credible frontier model, more than many expected after the weak reception of Meta’s models in 2025, and its lead in memory and long-running agent work gives Meta a distinct pitch for personal and professional agents. However, Meta has built its most capable model as a closed, hosted service, kept the strongest version behind a developer subscription, and asked buyers to trust it with personal data, and each of those choices works against it with the enterprise buyers it needs.
Our view is that Wonder’s benchmark position matters less than Meta’s route to market. Frontier models are now closely matched, and enterprises run several at once, so Meta does not need to beat Anthropic or Google outright to win workloads. It does need to be easy to buy, safe to connect to corporate data, and supported like an enterprise product, and on those measures, Meta has the most ground to make up of any frontier lab.
Memory and Long-Running Work Set Meta Muse Wonder Apart
Active Memory, Meta’s term for carrying context about a user across conversations and retrieving it when a task needs it, is the clearest technical difference between Wonder and its rivals. On MemoryBench, Meta’s own test of recall across long conversations seeded with distractors and facts that change partway through, Wonder scores 85.6% against roughly 81% for Opus 5.5 and Astra. Meta also reports leads on long-horizon coding, with 78.8% on DeepSWE 1.1 and 53.1% on SWE-Marathon against 46.3% for Opus 5.5, although rivals ran SWE-Marathon in their own agent harnesses, so that comparison is not controlled. Meta is plowing its own furrow here, building a model for an agent that knows its user rather than one tuned primarily for professional knowledge work.
On the Artificial Analysis Intelligence Index, Muse Wonder scores 52.4 at maximum reasoning effort, level with GPT-6 Astra (52.7) and just behind Claude Fable 5.1 (53.4), with Claude Sonnet 5.5 (56.0) and Claude Opus 5.5 (57.6) still ahead. Muse Wonder Full reaches 55.8, close to Anthropic’s models. Wonder’s strongest independent result is its hallucination rate, which Artificial Analysis measures at 26.5%, against 51% for GPT-6 Astra and 58.6% for Claude Opus 5.5. Every figure Meta published was run at maximum effort, however, and most of its agent results come from Meta’s own evaluations, so buyers should treat them as an upper bound.
Personal Data Is Meta’s Biggest Enterprise Risk
A personal agent that reads messages, calendars, and documents needs to handle private information well, and Meta’s own safety chart shows Wonder mishandling personal information in 3.4% of test cases. That is better than GPT-6 Astra and Fable 5.1 on Meta’s test; however, it still amounts to roughly one failure in 30. Futurum’s 1H 2026 AI Platforms Decision Maker Survey (N=820, February 2026) found that security and data privacy are the top concerns about agentic AI at every stage of deployment, cited by 21% to 28% of respondents, and unlike regulatory concerns, they do not fade with experience. For a company with Meta’s record on privacy, a one-in-30 error rate is a hard sell to enterprise risk committees.
Wonder’s model card states that Meta cannot rule out the unmitigated model reaching its high-risk threshold for chemical and biological misuse, so it shipped with classifiers and safety training, and the fuller safety report is still to come. Meta now also publishes a Meta Superintelligence Scaling Framework, joining the labs that use scaling and pacing commitments as much to signal competitive standing as to manage risk.
A Closed Model With a Subscription-Only Flagship
Meta’s decision to keep Wonder closed gives up one of the few enterprise advantages it built in the Llama era. Enterprises adopted Llama because they could run it in their own environments, fine-tune it on their own data, and avoid sending every prompt to a provider’s cloud. Meta still offers that with Muse Glimmer, a 30-billion-parameter model released under Apache 2.0 in August 2026, and it has said it will release the weights of Muse Spark 1.2, but its frontier model is now a hosted service governed by Meta’s commercial terms, like every other model at this tier. A frontier model that cannot be self-hosted puts Meta in direct competition with Google, OpenAI, and Anthropic on their terms, and those companies have years of enterprise sales, support, and compliance capability that Meta lacks.
Wonder Full, the model behind Meta’s strongest benchmark results, is available only through Muse Code plans and cannot be called through the API, so enterprises cannot build production applications on it. Wonder itself is sold mainly through Meta’s own API, with no confirmed presence yet on hyperscaler marketplaces, which our February 2026 survey suggests enterprises value, with 40.4% of model users accessing the same OpenAI models both directly and through Azure OpenAI. Meta AI chief Alexandr Wang said the API would be priced meaningfully below OpenAI and Anthropic, which holds against Fable 5.1 at $10/$50 but not against Opus 5.5, whose rate card matches Wonder’s $4/$20 down to the $0.20 cached input price.
How Muse Wonder Compares With Google’s Gemini 4 Argon
Gemini 4 Argon, which Google announced in September, is Wonder’s closest comparison, and the model Meta chose to leave out of its launch benchmarks. The two are level on general intelligence, with Argon at 52.6 on the Artificial Analysis index against 52.4 for Wonder, and both claim the lead on DeepSWE 1.1, with Google reporting 77.9% for Argon. Argon pulls ahead on legal work, where Google reports 19.6% on Harvey’s Legal Agent Benchmark against 10.0% for Wonder, on Terminal-Bench 4.0 (57.1% against 52.0%), and on hallucination, where Artificial Analysis measures Argon at 15% against Wonder’s 26.5%.
Futurum’s September 2026 AI Tools Pulse data (N=100) points to an even wider gap in distribution, with 61% of enterprise respondents reporting a business relationship with Google Gemini and 35% using it broadly across the company, while Meta is not among the four tools tracked. Argon will eventually underpin most of Google’s services and is likely to reach enterprises through Vertex AI and Google Workspace. Google is also releasing Argon in phases, starting with vetted cyber defenders in its Fairwind Program, so Wonder’s main advantage today is that buyers can use it now, although a public Argon rollout at Google’s introductory $2/$10 price would remove that head start.
Is Muse Wonder the Anchor for a Meta Cloud?
Wonder’s closed, API-only design makes more sense as part of a wider infrastructure strategy than as a model release on its own. Meta set up a Meta Compute division in January 2026 to run its data center build-out, and Bloomberg reported in July 2026 that Meta is building a cloud business to sell excess AI capacity to enterprises, either as raw GPU rental or as hosted models along the lines of Amazon Bedrock. Mark Zuckerberg has said a cloud business is definitely on the table. A hosted model service needs a model that buyers will pay for, and Muse Spark was not that model. Wonder, priced in line with Opus 5.5 and served through the Meta Model API, looks like the model layer of that cloud, and Meta has little reason to release the weights of the product it intends to sell.
Meta has guided to a 2026 capital expenditure of $115 billion to $135 billion, and selling capacity and model access is one of the few ways to earn a return on that spending outside advertising. However, Meta is also one of the largest buyers of third-party AI capacity, with reported contracts including more than $10 billion with Google Cloud, $14.2 billion with CoreWeave, and $3 billion with Nebius, so its first cloud customers are more likely to be AI labs buying overflow capacity than enterprises moving workloads. Its enterprise record is also thin, and Meta shut down Workplace, its enterprise collaboration product, in May 2026, which buyers will weigh before building on Meta infrastructure.
An Expanding Frontier Converges on Capability and Price
Meta’s arrival, alongside Argon and OpenAI’s GPT-6.1 Sol, shows how quickly the frontier is expanding and how alike its members have become. The six leading models are separated by just over five points on the Artificial Analysis Intelligence Index, every model at this tier is closed, and standard list prices are settling at $4/$20. Google and OpenAI launched at $2/$10 but plan to double those rates, so the September and October releases point to a common price for frontier capability rather than a price war. Output volume also varies widely at maximum reasoning effort, and Artificial Analysis found Argon used 110 million tokens to run its index against a median of 82 million, so cost per completed task is a better guide than list price, and no equivalent figure has yet been published for Wonder.
When capability and price converge, enterprises choose distribution, trust, and integration. Our February 2026 decision-maker survey found organizations run 3.8 models on average, so a new frontier model can win a place in a portfolio without displacing an incumbent. For enterprise buyers, our recommendation is to test Meta Muse Wonder for memory-dependent agents and long-running coding work, where its advantages appear strongest, and to budget at standard rates. Buyers should also expect Meta to keep Wonder off hyperscaler marketplaces if (and that’s a big ‘if’) it is building its own cloud, since listing it on AWS, Google Cloud, or Azure would strengthen the companies Meta plans to compete with. That leaves Meta choosing between the reach of the hyperscaler channels and control of its own, and a strong model that is hard to buy will struggle in a market this crowded.
What to Watch:
- Meta’s full safety report and independent testing of how Wonder handles personal data will determine whether enterprises trust Muse to operate inside email, calendars, and documents.
- Whether Meta lists Wonder on hyperscaler marketplaces or keeps it for its own cloud offering, and whether Wonder Full reaches the API, will show which route to enterprise procurement Meta has chosen.
- Uptake of the $20 and $100 Muse subscriptions once Wonder reaches the app will test Meta’s ability to earn AI revenue outside advertising.
- Whether Meta releases the Muse Spark 1.2 weights will indicate whether its open-weight line remains a serious enterprise option alongside its closed frontier models.
- A public rollout of Gemini 4 Argon, with a newer Gemini 4 version codenamed Carbon already in internal testing at Google, would remove Wonder’s head start on availability.
See the complete announcement of Muse Wonder and Muse Wonder Full on the Meta AI Research website.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Other Insights From Futurum:
Why Enterprise AI Budgets Keep Growing Despite AI Safety Pledges
Meta Reopens Its Models. Is This a PC Play or a Policy Play?
Google Returns to the Frontier With Gemini 4 Argon
OpenAI’s GPT-6 Astra: Benchmarks, Cyber Risks, and Market Impact
Author Information
Nick Patience is VP and Practice Lead for AI Platforms at The Futurum Group. Nick is a thought leader on AI development, deployment, and adoption - an area he has researched for 25 years. Before Futurum, Nick was a Managing Analyst with S&P Global Market Intelligence, responsible for 451 Research’s coverage of Data, AI, Analytics, Information Security, and Risk. Nick became part of S&P Global through its 2019 acquisition of 451 Research, a pioneering analyst firm that Nick co-founded in 1999. He is a sought-after speaker and advisor, known for his expertise in the drivers of AI adoption, industry use cases, and the infrastructure behind its development and deployment. Nick also spent three years as a product marketing lead at Recommind (now part of OpenText), a machine learning-driven eDiscovery software company. Nick is based in London.
