Meta ships Muse Wonder, its first frontier-tier model

Meta Superintelligence Labs (MSL) released a pair of models Muse Wonder and Muse Wonder Full  that it describes as its first models to reach the leading frontier tier. Where the earlier Muse Spark was tuned for low latency, the Wonder models are larger and, by Meta’s own testing, score comparably to Anthropic’s Fable 5.1 and OpenAI’s GPT-6 Astra. Access through the Meta Model API is priced at $4 per million input tokens and $20 per million output tokens  roughly the cost of OpenAI’s GPT-6 Sol or Anthropic’s Opus 5.5, and a clear step up from Spark’s lightweight tier.

Two models, two front doors

The release splits into two models with different distribution. Muse Wonder is available at launch through the Meta Model API and the Muse Code developer tool, and Meta says it will also power the company’s Muse agent reaching subscribers first along with artifact generation. Muse Wonder Full, the more capable of the pair, is restricted to Muse Code subscribers and is deliberately not offered through the API. Meta did not disclose parameter counts or context-window length for either model at launch.

Table 1. The Muse Wonder launch at a glance. Source: Meta AI Research blog and model card; TechCrunch.

Priced like a frontier model

The pricing is the clearest signal of where Meta is placing Wonder. At $4 and $20 per million tokens, it sits in the same bracket as the flagship tiers from OpenAI and Anthropic rather than the cheaper, high-volume tier Spark occupied. Meta’s cost comparisons, like its performance comparisons, are its own; independent cost-per-quality analysis will follow once third parties run the model.

Table 2. Meta Model API list pricing and Meta’s stated cost peers. Source: TechCrunch; Meta Model API.

The real pitch is memory

Rather than lead on raw intelligence scores, Meta built its messaging around two capabilities it calls Active Memory and Active Focus. Active Memory is the model carrying context about a user across conversations and pulling the relevant details back when a task needs them; Active Focus is sustained effort toward a goal trying different approaches, working through blockers, and checking in before taking consequential actions. To back the memory claim, Meta points to MemoryBench, a benchmark for continual learning and memory construction, and says Wonder now scores higher on it than Fable, Opus, or Astra. Research scientist Kavya Kopparapu said in a launch video that the team worked to make the model recall what matters to a user “over days, weeks, months, and beyond.”

The benchmarks Meta is leaning on

Beyond memory, Meta published results spanning agentic tasks, domain expertise in law, finance and health, focus, and coding. The headline numbers are all vendor-reported drawn from Meta’s model card and independent evaluators such as Artificial Analysis typically take weeks to post their own runs. The table below lists the benchmarks Meta cites and what each is meant to measure.

Table 3. Benchmarks Meta cites for the Wonder series. All figures are Meta’s own; exact scores sit in Meta’s model card and are not reproduced here because they are not yet independently verified.

A year of Muse, in context

Wonder lands about fourteen months after Alexandr Wang, the Scale AI founder Meta brought in to run MSL reorganized the company’s AI efforts in August 2025. In the months since, Meta shipped the lightweight Spark, media and voice models, an open-weight model (Glimmer), and a coding agent, but held back from the kind of top-tier release OpenAI and Anthropic ship routinely. At Spark’s April debut, Wang himself called it an “appetizer” and said it was below the tier of the leading frontier models  framing that pointed to exactly this release.

Table 4. Meta’s Muse releases in 2026. Dates per Meta’s research-blog index and launch coverage.

Compute and the safety framing

Meta says both models were trained at Prometheus, its gigawatt-scale data center in central Ohio that came online earlier this year. On safety, the company says Wonder carries its strongest safeguards to date against CBRN (chemical, biological, radiological and nuclear) and cyber misuse, while preserving frontier cyber capability for defensive use, under what it calls the Meta Superintelligence Scaling Framework. It calls Wonder its most aligned model so far and says a fuller safety report is coming.

What to watch

Three caveats stand out. First, the comparisons that matter most  parity with Fable 5.1 and GPT-6 Astra, and the MemoryBench lead are Meta’s own and await independent confirmation. Second, keeping Muse Wonder Full off the API limits who can test the stronger model. Third, the “first frontier model” label is contested: Spark was described as frontier-class when it launched in April, and Meta now reserves the top-tier framing for Wonder, a reminder that, in 2026, “frontier” is as much a positioning line as a technical one.