Meta's first frontier model, Muse Wonder, reaches the API at frontier prices

Meta released Muse Wonder and Muse Wonder Full on Monday, October 12. The company calls them its most capable models to date, and its newsroom post says they are built for "personal and professional agentic tasks." This is the first frontier-class release from Meta Superintelligence Labs since the company began rebuilding its AI organization roughly fourteen months ago, and it follows the lighter Muse Spark line.

The release matters less for the headline claim than for the fact that developers can actually buy it. Muse Wonder is available today through the Meta Model API and Muse Code, according to Meta. That puts a third vendor, alongside OpenAI and Anthropic, at the top of the price list for frontier API models.

What Meta actually shipped

• Muse Wonder: available through the Meta Model API and Muse Code starting today, per Meta's announcement.
• Muse Wonder Full: a larger variant. TechCrunch reports it will only be available to Muse Code subscribers and will not be served through the API.
• Muse agent: Meta says Wonder will "soon" power more of its Muse agent, starting with subscribers. The Verge notes Meta gave no timeline.

Meta highlights two behaviors. The first it calls Active Memory: carrying what matters to a user across conversations and bringing it back when relevant. The second it calls Active Focus: staying on a goal when a task gets hard, trying other approaches, recovering from errors, and checking in before doing anything important. The company names law, finance and health as target domains, along with coding.

Price and performance: Meta's numbers, not yet anyone else's

According to TechCrunch, which cites Meta's model card, Muse Wonder costs $4 per million input tokens and $20 per million output tokens on the Meta Model API. TechCrunch describes that as roughly the same as OpenAI's GPT-6 Sol or Anthropic's Opus 5.5. Promptea could not open Meta's model card or API documentation from our research environment, so these prices come from TechCrunch's reading of the card rather than from a page we checked ourselves.

On capability, Meta says its internal testing puts Wonder and Wonder Full on par with frontier models from OpenAI and Anthropic. TechCrunch names Fable 5.1 and GPT-6 Astra as the comparisons; The Verge lists GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Meta also says Muse Wonder beats Fable, Opus and Astra on MemoryBench, a benchmark for continual learning and memory construction. All of these results come from Meta's own testing. Independent evaluation is not yet available. The Verge also points out that it is unclear how the models compare with Google's Gemini 4 Argon.

On safety, Meta says the models were trained to avoid sharing private information without permission and to refuse genuinely harmful requests, while still supporting defensive security work such as finding flaws in your own code. It says it followed its Meta Superintelligence Scaling Framework and that "a full report with additional safety evaluations will be available soon." That report was not out at launch, which matters for a model sold specifically for agentic work.

What this changes for developers

Model choice. If you are already paying frontier prices for agentic or coding workloads, there is now a third option at roughly the same price point, according to TechCrunch's comparison. The sensible way to evaluate it is to run your own tasks through it. A model that is "comparable" on a vendor's selected benchmarks can behave very differently on a specific codebase or document set.

The strongest variant is not in the API. Wonder Full is limited to Muse Code subscribers, according to TechCrunch. Anyone building a product on the API gets Wonder, not Wonder Full, so benchmark figures quoted for Wonder Full do not describe what you can integrate.

Memory is the claim to test. Active Memory and the MemoryBench result are what Meta leans on most. Meta's announcement does not say whether, or how, Active Memory is exposed through the API, as opposed to inside Meta's own Muse products. Until the documentation says, assume you still manage context yourself: what goes in the prompt, what gets summarized, and what is retrieved per turn.

What to watch

• Independent benchmark results for Muse Wonder, especially on coding and long-horizon agent tasks.
• Meta's promised safety report and the full model card.
• API documentation clarifying how Active Memory and Active Focus work for developers, if at all.
• Whether Wonder Full ever reaches the API.