TL;DR: Meta released Muse Wonder and Muse Wonder Full on Monday, its first models that genuinely compete at the top of the AI leaderboards. Independent testing puts Wonder Full level with Anthropic’s Claude Sonnet 5.5 and just behind Claude Opus 5.5, but the stronger model is only available inside a paid Muse Code subscription, not through Meta’s API.
For most of the last two years, Meta’s AI story has been about distribution, not cutting-edge models. It had billions of users and a lot of data centers, but the most capable systems kept coming from OpenAI, Anthropic and Google. On Monday, October 12, that changed, at least on paper. Meta Superintelligence Labs, the unit run by former Scale AI CEO Alexandr Wang, released two new models, Muse Wonder and Muse Wonder Full, which TechCrunch describes as the company’s first real attempt at a frontier model.
“Muse Spark was our first family of models, and now Muse Wonder is a much larger and more advanced class of models that brings us to the frontier,” Mark Zuckerberg wrote on X, as reported by AFP. Wang went further in an interview with Axios, calling them “true frontier models, competitive with the best models in the world.” Internally, the models were code-named “Watermelon.”
What Muse Wonder is built to do
Meta isn’t pitching Wonder as a better chatbot. It’s pitching it as an engine for AI agents that work across apps for long stretches without losing track of the goal. The company is leaning on two capabilities it calls Active Memory and Active Focus. According to VentureBeat’s breakdown, Active Memory lets an agent hold on to personal and professional context across conversations, while Active Focus is meant to keep it pursuing an objective when it hits obstacles, changing requirements or failed attempts, including asking permission before consequential actions.
Both models accept text, images, audio and video, respond in text, and support context windows of roughly one million tokens. Meta says they were trained in its Prometheus data center, combining large-scale pretraining with reinforcement learning. In its own demos, Meta shows agents building software, organizing complicated personal plans and handling work like financial analysis and legal document prep.
How it stacks up in independent tests
Benchmarks published by the company that made the model always deserve a raised eyebrow, so the more useful numbers come from Artificial Analysis, which tested both models itself. On its Intelligence Index, a composite of ten evaluations, Wonder Full scored 56 and the standard Wonder scored 52, up from 48 for Meta’s previous Muse Spark 1.3. Artificial Analysis notes the jump came only six weeks after Spark 1.3.
For context, VentureBeat reports that 56 ties Claude Sonnet 5.5, beats GPT-6 Astra, Gemini 4 Argon and Claude Fable 5.1 (all at 53), and trails only Claude Opus 5.5 at 58 among the models it compared. The standard Wonder sits one point behind GPT-6 Astra.
A few details from the Artificial Analysis write-up stand out:
- Agentic work is the big gain. Wonder Full’s score on Terminal-Bench 4.0 jumped from 33% to 57% compared with Spark 1.3, and it ranks near the top on several “knowledge work” evaluations.
- Coding agents got much better. Running inside Muse Code, Wonder Full scored 68 on the firm’s Coding Agents Index, level with Claude Code running Sonnet 5.5 and ahead of Opus 5.5 at 66. Spark 1.3 scored 54.
- It hallucinates less. Wonder Full’s hallucination rate fell from 33% to 23%, the second lowest among models scoring 45 or more on the index, behind only Gemini 4 Argon.
- Raw knowledge is still a weak spot. On the AA-Omniscience knowledge test, both models trail Opus 5.5, GPT-6 Astra and Gemini 4 Argon.
Meta’s own model card adds more flattering comparisons. On SWE-Marathon, a long-running coding benchmark, Meta reports Wonder Full at 57.5 versus 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra. But it doesn’t win everywhere: Opus 5.5 and GPT-6 Astra both edge it on Terminal-Bench 4.0 in Meta’s table, and Opus leads on the HealthBench Pro health questions. VentureBeat also points out that Meta’s comparison mixes its own evaluations, vendor-supplied scores and third-party results, so it’s less controlled than testing everything under identical conditions.
Pricing, and the catch for developers
Through the Meta Model API, the standard Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million. That’s the same per-token price as Claude Opus 5.5 and well below GPT-6 Astra. Web search through the API is billed separately at $2.50 per 1,000 queries.
Per-token price isn’t the whole story, though. Artificial Analysis estimates Wonder costs $4.52 to run an average task on its index, because it chews through a lot of input tokens. That’s cheaper than Opus 5.5 at $5.98 but pricier than GPT-6 Astra at $3.26, and nearly three times what Muse Spark 1.3 cost.
The bigger catch is that Wonder Full, the stronger model, isn’t on the API at all. It’s only available through Muse Code, Meta’s coding assistant, on a $50-a-month “Power Usage” plan or a $200-a-month “Maximum Usage” plan, according to VentureBeat. That’s fine for developers who want to try it, but it means businesses can’t yet build Wonder Full into their own products.
There’s also a data question worth reading twice. Meta says content sent through the pay-as-you-go API, which runs in what it calls Standard mode, isn’t used for training. Muse Code subscriptions, however, default to “Contributor mode,” which lets Meta use prompts, code and files to improve future models. Subscribers can switch it off, but only for future submissions.
A break from Meta’s open-model past
It’s also notable what Meta didn’t do. Its Llama models helped popularize AI that anyone could download and run on their own hardware. Neither Wonder model is being released with downloadable weights. Both are hosted, proprietary services.
The release caps a costly rebuild. As The Business Times notes, Zuckerberg spent the last year and a half refocusing Meta on the AI race, committing hundreds of billions of dollars to infrastructure and bringing in Wang as part of a $14 billion investment in Scale AI. Investors seemed to like the result: Meta shares rose 3.4% on Monday, according to Reuters’ market wrap.
Wang also told reporters Wonder is larger and more expensive to run than Spark, so Meta is already working on specialized versions that keep much of its capability while being cheap enough to serve at Meta’s scale. Wonder will also power the subscription version of Meta’s Muse personal agent app.
The open question, as Axios puts it, is whether Meta’s claimed performance holds up in independent testing over time, and how quickly any of this reaches the billions of people who use Meta’s apps.
Related: Nvidia Is Reportedly in Talks to Buy the Open-Source AI Lab It Already Backs for $25 Billion and Google and Unity Will Let You Build a Video Game By Just Typing What You Want.
Bottom Line: Muse Wonder is the first time Meta has shipped a model that independent testers place alongside the best from Anthropic and OpenAI. The standard model is available now at Opus-level prices, but the version that actually reaches the frontier is locked inside Muse Code for the moment, and it’s no longer open-weight.
Follow Teck Hustlers: Facebook · X · Instagram · LinkedIn · TikTok · Pinterest
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
