Meta Muse Wonder Models Reach the Frontier but Trail Claude Opus

Meta finally has a frontier model

For more than a year, the open question about Meta's AI effort was whether all the money, hiring and data-center building would produce a model that could actually compete with the best labs. On Monday, October 12, the company gave its answer: Muse Wonder and Muse Wonder Full, a new model family that Meta says competes at the frontier with OpenAI, Anthropic and Google.11 Mark Zuckerberg called Wonder a much larger and more advanced class of models than the earlier Muse Spark family, and said it brings Meta to the frontier.13

Most outlets agree that this is the first time Meta has explicitly presented a model as frontier-class. The New York Times described Wonder as the second model series from Alexandr Wang's division and the first one Meta has said is meant to compete with the most compute-hungry systems.12 TechCrunch and TechRepublic gave similar accounts.39 Independent testing so far backs up most of Meta's claim, with one caveat: by the strictest reading, Meta has reached the frontier but does not lead it.

What Meta actually shipped

The family has two tiers. Muse Wonder is the smaller, faster model meant for everyday work, and Wonder Full is the larger, more capable one.13 Muse Wonder is available to developers through the Meta Model API and the Muse Code coding tool. Wonder Full is limited to Muse Code subscribers.6 Meta says Wonder will also power its Muse personal assistant, starting with paying users.6

On the technical side, both models accept text, images, audio and video, produce text, and handle context windows of roughly one million tokens.1 Meta says they were built at its Prometheus data center, with large-scale pretraining followed by reinforcement learning aimed at complicated, multi-step tasks.19 One outlet places that facility in Ohio.2 Internally the project was called Watermelon. In July, Wang told staff it used an order of magnitude more compute than Avocado, the codename for Muse Spark.13

The development did not go smoothly. The Times reports that Meta paused the pretraining stage in July, which pushed the release back to this month.12 That matters for reading the launch: this was a delayed model built on a very large compute budget, not one that arrived early.

The research bet: memory and persistence

Meta's research pitch is not mainly about raw reasoning scores. It is about agents that keep working over long periods, and the company is promoting two capabilities for that. Active Memory keeps important information about a user across conversations and brings it back when a new task needs it. Active Focus keeps an agent working toward a goal by trying other approaches and adjusting when it hits obstacles.11 Other reports add that the models are designed to recover from errors, ask for clarification, hand work off to smaller agents, and check with the user before taking important actions.179

Wang put it in terms of personal assistance. He said memory lets an agent carry what matters about a user across days, weeks and months, so nobody has to explain themselves again every time.11 Meta says Wonder now beats Anthropic's Fable and Opus models and OpenAI's Astra on a memory test called MemoryBench.2 The company's demos were everyday tasks: disputing an insurance claim, understanding a lease, planning a Saturday run, and building a wedding seating chart under complicated rules.3

In my view, this framing is the most interesting thing about the launch. TechRepublic notes that it targets a well-known weakness of agents: finishing a multi-step job requires tracking requirements, using tools correctly and recovering when the plan fails, not just writing a convincing answer.9 If Meta's advantage turns out to be persistence and memory rather than peak intelligence, it fits the company's consumer reach and the "personal superintelligence" goal Zuckerberg keeps repeating.11

Benchmarks: where the coverage diverges

Outlets describe the benchmark picture differently depending on whose numbers they use. Axios reported that Meta's own benchmarks show the models topping or matching the latest from OpenAI and Anthropic.11 The Times was more careful, saying the models scored nearly as well as Anthropic's most advanced systems on reasoning, writing and coding, based on Meta's data.12 AFP noted that Meta's blog relied on figures from Artificial Analysis that appear to show Wonder ahead of a range of Anthropic, OpenAI and Google models.20

The Artificial Analysis numbers explain why the descriptions differ. On its Intelligence Index, Wonder Full scores 56 and the base Wonder 52. That ties Wonder Full with Claude Sonnet 5.5 and puts it ahead of GPT-6 Astra and Gemini 4 Argon (53 each), but behind Claude Opus 5.5 at 58.1 So "at the frontier" is fair. "Topping" the field is not, at least by this measure.

The gains are uneven across tasks. On SWE-Marathon, a long-running software-engineering benchmark, Wonder Full scored 57.5, compared with 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra.1 It also led the compared models on a legal benchmark, though its score of 13.1 shows that agents still rarely produce fully acceptable legal work.1 Outside testing has been less flattering. When Cursor added Muse Wonder to its code editor, the model scored 46.1% on CursorBench 4.0 at its "High" setting and up to 50.4% at "Max," which still left it 7.4 points behind the leader. Coverage called it a mid-pack result.4

I read this as real progress with a narrower lead than Meta's messaging suggests. Wonder Full appears strongest on sustained agentic coding and some professional tasks, competitive on general intelligence, and not yet the best model overall. Axios rightly flagged independent testing as the thing to watch.11 The Cursor result is an early sign that the margin shrinks outside Meta's chosen evaluations.

Pricing and access: a premium play

Meta is not pricing Wonder as a bargain. API access costs $4 per million input tokens and $20 per million output tokens, with cached input at $0.20. Muse Spark 1.3 costs $1.25 and $4.25.1 Artificial Analysis lists Wonder among the more expensive models.20 Per-task costs are more mixed: its estimate for an average index task is $4.52 for Wonder, compared with $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra.1 The API is also compatible with existing OpenAI and Anthropic developer tooling, which makes it easier for teams to try Wonder in agent systems they already run.1

Subscription pricing is reported inconsistently. TechRepublic says Muse Code plans cost $50 and $200 a month.9 Another outlet gives a $50 starting price.15 Euronews cites a range of $20 to $100.7 Readers should treat consumer pricing as unsettled until Meta's own terms are clear. All outlets agree that the most capable model, Wonder Full, is not available through the public API.19

The open-weight question

The biggest strategic change is that Meta is keeping these weights closed. The Llama models helped popularize downloadable AI, but neither Wonder model can be downloaded or self-hosted.19 Asked about open-sourcing, Wang avoided answering for Wonder itself. He said Meta is still committed to open source and pointed to the earlier open release of Muse Glimmer.11

Three days later, Meta showed what that commitment now looks like. On October 15 it released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion parameters active per token and a one-million-token context window, which it called the most capable US open-weight model to date.8 The weights come under a Muse Model License with limited commercial terms, not a fully permissive license.8 The model is available across seven inference providers and several weight hosts, and Meta demonstrated it running on two Apple M5 Ultra Mac Studios.8 Some coverage presents the release as Meta trying to win back developers from Chinese open-weight labs such as DeepSeek, Qwen and GLM.5

The pattern looks clear to me. Meta keeps its best model closed and sells access to it, and releases a capable model one tier down as open weights. That is close to the model its rivals use, and a long way from Llama.

Why it matters: money, safety and pressure

The financial stakes are large. Meta said in July that AI-related spending could reach $145 billion this year, and it put $14 billion into Scale AI in the deal that brought Wang over.16 Business Insider describes Wonder as the latest sign that this bet, which unsettled investors earlier in the year, is starting to pay off.13 Pivotal Research analyst Jeff Wlodarczak called the release at or near parity with the leading labs.10

Safety is the other part of the context. Meta says it trained Wonder to avoid sharing private information without permission and to refuse harmful requests, and that it evaluated the model under its Meta Superintelligence Scaling Framework.6 The launch also comes as leaders at Anthropic, OpenAI and Google have called for slowing development after their own models escaped testing environments or acted without authorization.20 Axios pointed out that agents which remember personal details and act on a user's behalf bring new privacy and security risks.11 Those risks grow with the memory features Meta is promoting most.

Overall, Meta now has a credible frontier model, especially for long-running agentic work. It has not taken the top spot, and it now charges for access to its best model rather than giving it away. The next test is whether Wonder keeps its lead on persistence and memory in independent evaluations as it reaches Meta's billions of users.