Two versions, two ways to buy
On October 12 Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full. Meta's earlier Muse Spark was built for low latency; this is its first frontier model to compete head-on with OpenAI's and Anthropic's flagships, trained at its Prometheus data center in Ohio.
The two versions are sold differently:
- Muse Wonder: available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, $0.20 for cached input, about 1 million tokens of context;
- Muse Wonder Full: no API; available only to subscribers of the coding product Muse Code, on $50 and $200 monthly plans.
Neither has open weights; Meta's open-weight line remains the smaller Muse Glimmer. The Muse app will move to the new models "soon," starting with subscribers.
Memory is the pitch, long-horizon coding is the result
Meta highlights two capabilities: "Active Memory," remembering and synthesizing information users provide across conversations, and "Active Focus," pushing toward a goal over long stretches and checking with the user before important actions. Research scientist Kavya Kopparapu said in the launch video that the team worked hard to make the model remember "what's important to you" over days to months.
The standout result is SWE-Marathon, an ultra-long software project benchmark: by Meta's figures as compiled by VentureBeat, Wonder Full scores 57.5 and Wonder 53.1, against 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra; on DeepSWE 1.1 Wonder scores 78.8, also above Opus 5.5's 74.2. But it isn't ahead across the board: of the 8 comparisons VentureBeat counted, the standard model beats Opus 5.5 on only 3, trailing on Terminal-Bench 4.0, HealthBench Pro and others. Rival models ran in their own agent harnesses, so comparisons like SWE-Marathon aren't strictly controlled.
Choosing by price
The standard model is priced exactly like Opus 5.5 and trades wins with it, so whether to switch depends on how much your work leans toward very long coding tasks; compare actual token usage and bills on your own tasks. The Full version, which leads by more, is locked inside Muse Code subscriptions, so developers who want to call it from their own products can't yet. Meta says a full safety report is "coming soon."