Meta Muse Wonder AI models put Meta at the frontier, near Claude

Meta's first serious bid for the frontier

On Monday, Meta released Muse Wonder and Muse Wonder Full. They are the first models from Meta Superintelligence Labs that the company says can match the top systems from OpenAI, Anthropic and Google.1 Inside Meta the family went by the code name "Watermelon," and the company says it was trained at Prometheus, its gigawatt-scale data center.1 TechCrunch describes the release as Meta's first real attempt at a frontier model. It follows Muse Spark, which was built to be lightweight and fast.9

Mark Zuckerberg presented the launch as a step up in size and ambition. He wrote that Muse Spark was the company's first model family and that Wonder is a much larger class of models that "brings us to the frontier."2 Alexandr Wang, who runs the lab, went further. He told reporters these are "true frontier models" and that they lead the field on personal and professional agentic work.18

My reading of the coverage: Meta now has a model in the top tier for the first time in years, but Anthropic still holds the top spot. Meta's strongest case rests on the specific tasks it picked to highlight, not on the broad measures that rank models overall.

What the benchmarks actually show

Reports agree on the broad shape and differ on how generous they are. Meta's own chart shows the models matching or beating OpenAI and Anthropic across tasks involving agents, memory and coding.1 The New York Times was more cautious, saying the models scored nearly as well as Anthropic's best on reasoning, writing and coding, according to Meta's data.6 TechCrunch reported that Meta's model card puts Wonder and Wonder Full roughly level with Claude Fable 5.1 and GPT-6 Astra.9

The independent numbers help settle it. On the Artificial Analysis Intelligence Index, which combines ten tests, Claude Opus 5.5 scores 57.6, Muse Wonder Full 55.8, Claude Fable 5.1 53.4, GPT-6 Astra 52.7 and Gemini 4 Argon 52.6.12 The standard Muse Wonder scores 52.4, a little behind OpenAI's and Google's flagships.12 One analysis noted that the index scores Meta printed on its own chart match the independent results exactly.13 That matters because it suggests Meta did not inflate the comparison it can least control.

So the headline result holds up: the biggest Wonder model sits in second place, behind Opus 5.5 and ahead of everything else on that index.1012 Some coverage stretched that into Meta "overtaking" GPT-6 and Gemini 4. That is accurate only for the Full variant at its highest compute setting.12

The task-level picture is more mixed. In a breakdown of Meta's 14-row comparison table, Wonder Full wins seven rows, and two of those wins come from benchmarks Meta built itself: its Personal Superintelligence Index and MemoryBench.13 The clearest leads are in long coding projects and legal agent work. On SWE-Marathon, Wonder Full scores 57.5 against 46.3 for Opus 5.5 and 42.5 for GPT-6 Astra.13 On DeepSWE 1.1 it scores 79.0, compared with 74.2 for Opus 5.5.13 It trails on ProgramBench, where Opus 5.5 scores 70.5 to Wonder Full's 59.0. It also trails on HealthBench Pro, Terminal-Bench and OSWorld computer-use tests.13

In other words, Meta built a model that does unusually well on long, multi-step tasks, which is the kind of work it says personal agents need. It is not yet the best general-purpose model.

Memory and persistence as the research bet

The main technical idea in the launch is two features Meta calls Active Memory and Active Focus. Active Memory lets the model keep relevant facts about a user across conversations and bring them back when needed. Active Focus is meant to keep an agent working toward its goal, trying other approaches when it gets stuck.1 Wang said the aim is for an agent to remember what matters to a person over "days, weeks, months" so users don't have to explain themselves again.1

Wired highlighted a safety-related side of Active Focus: keeping a model fixed on its stated goal could cut down on the drift and compounding errors that make long agent runs expensive.10 Meta pointed to MemoryBench, a test of continual learning and memory construction, as evidence. Wonder beats Fable, Opus and Astra there.9 But MemoryBench is one of Meta's internal benchmarks, and Wonder's lead over Opus 5.5 is about four points (85.6 to 81.5).13 Outside testing of memory will matter more than any other number in this launch, because memory is the capability Meta is building its strategy around.

Meta also says the models do especially well on unfamiliar tasks, which would mean their abilities reach beyond what they were trained on.1 For now, that is a company claim waiting for independent checks.

The pattern isn't new. Last month's Muse Spark 1.3 showed a similar profile: strong results on long-horizon coding benchmarks such as DeepSWE, while Opus 5 still led on broader knowledge-work and computer-use tests. Wonder looks like a scaled-up version of that recipe rather than a change of direction.

Price, access and the open-source question

Coverage of pricing doesn't agree. Through the Meta Model API, Wonder costs $4 per million input tokens and $20 per million output tokens, which TechCrunch puts at roughly the same level as Opus 5.5 or GPT-6 Sol.9 Wang told reporters the API would cost "significantly lower" than leading OpenAI and Anthropic models.3 The published rates don't clearly support that. One analysis pointed out that at the same per-token price, Opus 5.5 scores six points higher on the independent index than the API version of Wonder.13 That makes Meta's value argument weaker for developers choosing purely on capability per dollar.

Access comes in tiers. Wonder is available through the API and Muse Code. Wonder Full is limited to Muse Code subscribers, at $50 or $200 a month, and is not offered through the API.39 Wonder will also power Meta's Muse assistant, starting with paid subscribers.1

Reports also differ on what this means for Meta's open-source past. Stocktwits presented the closed, paid model as a break from Meta's Llama-era approach.4 Axios reported that Meta has not said whether any Wonder model will get open weights. Wang said the company is "still committed to open source" and pointed to the open-weight Muse Glimmer.1 Glimmer is a 30-billion-parameter model released under Apache 2.0 in August, and it scored 35 on the Artificial Analysis index.1117 That score is far below Wonder's. The practical result is a split: Meta gives away small models and charges for the frontier ones. That is a big shift from the Llama years, when Wired notes Meta shipped the first frontier open-weight models from a major lab.10

The cost of getting here

This launch tests one of the most expensive rebuilds in the industry. Meta hired Wang in June 2025 as part of a $14 billion-plus deal involving Scale AI, after Llama 4 disappointed.4611 Wired notes the effort has sometimes been rocky, with some researchers leaving quickly.10 The Times reported that Meta paused pre-training in July, which pushed the release back to this month.6 In May, Meta cut about 10% of its staff, roughly 8,000 jobs, partly to pay for infrastructure spending.4

Zuckerberg said reaching this point took both heavy compute spending and "very rigorous scientific work," and predicted that model development would speed up from here.4 Investors seemed persuaded on the day: Meta shares rose about 4% after the announcement.34

A different destination than its rivals

Meta is pitching Wonder as a step toward "personal superintelligence," not artificial general intelligence.1 Zuckerberg contrasted this with competitors that he said focus on automating knowledge work.4 Wang made a similar point, saying the rest of the industry has concentrated on coding agents.8 There is some tension here, because Wonder's clearest benchmark wins are in coding and legal work, and Meta is openly chasing enterprise customers.313

The consumer side gives the strategy some weight. Muse, the agent app launched September 8, hit 5 million downloads in 22 days, faster than ChatGPT, Grok or Claude according to Sensor Tower.8 Meta's distribution across 3.5 billion users is still an advantage no pure AI lab has.6

Safety and what to watch

Meta says it has tightened safeguards against cyber and biological misuse and published a model card, with a fuller report to come.1 Wired noted that Meta did not say whether Wonder-powered agents misbehaved during testing. That gap stands out after a run of incidents involving rogue agents from other labs.10 Zuckerberg rejected calls for an industry-wide slowdown, arguing that AI can be built with "normal guardrails."4

The open questions are clear. Independent memory and long-horizon evaluations need to confirm Meta's internal numbers. Wang has told skeptics to try the models and judge for themselves.3 It also remains to be seen how quickly Wonder reaches Meta's billions of app users.1 For now, Meta has shown that its rebuilt lab can reach the frontier. It has not shown that it can lead it.