Meta Muse Wonder launches as its first frontier model

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, the first models it calls frontier class. A frontier model is a lab's largest, most capable system. The smaller Muse Wonder is open to developers today at $4 per million input tokens and $20 per million output tokens. The larger Wonder Full stays inside Meta's own coding product.

That split matters for anyone choosing a model. The version that tops most of Meta's own charts is the one you cannot call from code.

Two models, two ways in

Meta's announcement calls the pair "our most capable models to date." Both are built on Muse Spark, the smaller model family Meta launched in April 2026. Meta says Muse Wonder "will soon power more of Muse," its personal agent app, rolling out first to paying subscribers.

TechCrunch reports that both models came from Meta's Superintelligence Lab. They were trained at Prometheus, Meta's data center in central Ohio. It is the lab's first frontier release since Alexandr Wang restructured Meta's AI work in August 2025.

	Muse Wonder	Muse Wonder Full
Meta Model API	Yes, $4 in / $20 out per million tokens	No
Muse Code	All plans, from $5 a month	Power ($50) and Maximum ($200) plans
Context window	1 million tokens	1 million tokens
Weights	Closed	Closed

The plan prices, context window and closed weights come from FelloAI's launch summary. A token is a small chunk of text, roughly three quarters of a word. The context window is how much text a model can read at once.

FelloAI also reports a cached-input price of $0.20 per million tokens. That is the discounted rate for text the model has already seen. It puts Muse Wonder at about 3.2 times the input price of Muse Spark 1.3, and 4.7 times the output price.

What the benchmarks show

Meta's model card says the two models score "comparable" to Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra, according to TechCrunch. On MemoryBench, a test of what a model remembers across conversations, Meta says Muse Wonder beats Fable, Opus and Astra.

FelloAI compiled Meta's published scores. A few show where each model leads:

Benchmark (Meta-reported)	Wonder Full	Wonder	Claude Opus 5.5	GPT-6 Astra
MemoryBench	85.5	85.6	81.5	81.4
DeepSWE 1.1 (long coding tasks)	79.0	78.8	74.2	74.1
Terminal-Bench 4.0	57.1	52.0	59.6	59.1
ProgramBench (code from scratch)	59.0	50.0	70.5	52.5
OSWorld V2.1 (computer use)	78.4	72.6	81.8	75.8

The independent Artificial Analysis index tells a cooler story. It ranks Muse Wonder at 52, below Claude Opus 5.5 at 58 and one point behind Fable and Astra at 53. Wonder Full scores 55.8, but that model has no public API.

Memory and safety claims

Meta's post highlights Active Memory, which it describes as "the ability to carry what matters to you across conversations." It also says the models were trained "to avoid sharing private information without permission."

The safety picture is incomplete. Meta says a full report with more safety evaluations "will be available soon." FelloAI's reading of the model card says Wonder Full may reach a high-risk level for chemical and biological misuse without safeguards. The card puts the remaining risk at "moderate or lower" once safeguards are in place.

The card also reports an ExploitGym test, in which Wonder Full flagged 92 security issues. Of those, 21 came through the intended vulnerability and 71 through holes the test designers had not planned. Meta says the models are not for users under 18.

What this means for developers

Muse Wonder enters at the same price FelloAI lists for Claude Opus 5.5: $4 in, $20 out and $0.20 for cached input. That makes the comparison direct. Run both on the same tasks before switching anything.

• Test what you can actually call. Most of the strongest numbers belong to Wonder Full, which is only in Muse Code. Benchmark Muse Wonder itself, the API model, on your own workload.
• Try it on long, memory-heavy work. Meta's clearest leads are MemoryBench and long coding tasks such as DeepSWE. Short, single-shot coding is where Opus 5.5 still scores higher on Meta's own chart.
• Recheck your budget if you used Muse Spark. Moving from Muse Spark 1.3 to Muse Wonder roughly triples input costs and nearly quintuples output costs.
• Do not plan around self-hosting. Both models are closed. FelloAI notes that the open Muse Spark weights Meta promised for September 2 have still not shipped.
• Wait for the safety report before giving the model broad tool access in production. Meta has said a fuller report is coming but has not dated it.

The release continues a fast run of Muse updates, from the Muse Spark 1.3 coding update in September to the Muse Gadgets hardware kits on October 2.