A frontier release at last
Meta released Muse Wonder and Muse Wonder Full on Monday, its first models aimed at the frontier rather than at low cost and low latency. The launch post from Meta’s Superintelligence Lab calls the pair “the next step on our scaling ladder and a major leap toward personal superintelligence”, and says both were trained at Prometheus, the gigawatt-scale data centre Meta brought online in central Ohio this year.
Both take text, images, video and audio and return text, with a one-million-token context window at inference, according to Meta’s model card. Meta has given them two named behaviours. Active Memory carries what the model judges important about a user from one conversation to the next. Active Focus keeps working at a blocked task by trying other routes, and checks in before taking an important action.
Fourth on an index Meta itself cites
Artificial Analysis scores Muse Wonder Full at 56 on version 4.3 of its Intelligence Index, fourth of the 689 models it has measured. The smaller Muse Wonder scores 52, and Muse Spark 1.3, the model it succeeds, scored 48. Meta’s own scorecard uses the same index, so for once the vendor’s chosen headline figure and the independent measurement are the same number.
Vals AI, which grades against a private test set, places Muse Wonder Full fifth on its own index and ahead of every OpenAI model, with only Anthropic’s current releases above it. Muse Wonder lands tenth.
Where Meta reports its own comparisons, they are harder to read. The model card gives Muse Wonder Full 13.1% on Harvey’s LAB-AA legal-agent benchmark — a figure it attributes to Artificial Analysis — against 4.2% for Claude Opus 5.5 and 8.6% for GPT-6 Astra; the measure credits only deliverables that meet every criterion and pass a hallucination check. On SWE-Marathon, where the card names Abundant AI as the result source for all models, Meta reports 57.5% against 46.3% and 42.5%, but adds that the row is “not a controlled common-harness comparison” and that Muse Wonder Full ran at its highest reasoning setting. MemoryBench, where Meta reports 85.5% against 81.5% for Opus 5.5, is Meta’s own benchmark and is graded by a language model.
The stronger model is not for sale
Muse Wonder is available now through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, against $1.25 and $4.25 for Muse Spark 1.3. The Decoder, citing Artificial Analysis, puts the cost of one task on the Intelligence Index at $4.52 for Muse Wonder against $1.60 for its predecessor.
Muse Wonder Full is restricted to Muse Code subscribers and is not served through the API at all. Artificial Analysis records “no publicly available API” for it, which is why the model Meta ranks fourth carries no price and no speed figure. Meta says Muse Wonder will power the Muse app soon, beginning with paying subscribers, and that it plans to build specialised models from it for everyone over time.
What the card concedes
Meta lists what the models do not do well. Visual and spatial reasoning, “especially 3D tasks, remains a known limitation”, and Meta tells developers to validate performance themselves wherever precise spatial understanding matters. The card also says the models are not intended for anyone under 18, and puts the burden of assessing any such use on whoever deploys them.
Kavya Kopparapu, a research scientist at Meta, said in the launch video that the team “put a tremendous amount of effort into making sure your model can remember what’s important to you”, according to TechCrunch.
This is Meta’s first frontier-tier release since Alexandr Wang reorganised the company’s AI work in August 2025, a stretch in which it shipped the lightweight Muse Spark and the Muse agent but stayed out of the state-of-the-art race. The question now is whether Muse Wonder Full reaches the API, and at what price. For the moment, the model Meta sells is the weaker of the two.
