Meta releases Muse Wonder models with gains in coding and agent tasks
Meta has released Muse Wonder and Muse Wonder Full, larger AI models aimed at complex tasks such as coding and personal assistance. Reported benchmarks show stronger performance than their predecessor, alongside higher costs and continuing weaknesses in factual knowledge.
wired.comMuse Wonder Puts Meta Back at the AI Frontier
techcrunch.comMeta releases its first frontier model, Muse Wonder
the-decoder.comMeta's new frontier model Muse Wonder Full breaks into the AI top tier
Two new models
Meta released Muse Wonder and Muse Wonder Full on Monday, expanding its Muse lineup with larger models designed for more complex tasks than the lightweight Muse Spark. Meta describes the new models as a step toward its goal of personal superintelligence.
Muse Wonder is available through Meta’s API and its Muse Code software-writing tool. TechCrunch and The Decoder report that Wonder Full is currently restricted to Muse Code users, although Wired describes the models collectively as available through both channels.
Performance gains and limits
The Decoder reports that Muse Wonder Full scores 56 on the Artificial Analysis Intelligence Index, compared with 52 for Muse Wonder and 48 for their predecessor, Muse Spark 1.3. Wonder Full ranks in the top three on six of ten benchmarks, with particularly strong gains in tasks that require models to work independently.
Coding results also improved: Wonder Full scores 68 on the Coding Agents Index, while Wonder scores 63, up from 54 for its predecessor. But factual knowledge remains weaker: Wonder Full scores 30 on the AA-Omniscience test, behind Claude Opus 5.5 at 46 and GPT-6 Astra at 43.
Meta says its largest model outperforms leading rivals on some specialized benchmarks. Wired cautions that benchmark scores do not always reliably measure real-world performance.
Memory, focus and safety
Meta highlights two features: Active Memory, designed to retain important information across conversations, and Active Focus, intended to help models persist with difficult tasks, try alternative approaches and ask clarifying questions. Both support the company’s ambitions for personal AI agents.
Meta says it trained the models to refuse dangerous requests, including assistance with bioweapons or hacking. Wired reports that the company did not disclose whether Muse Wonder agents had misbehaved during testing or what specific measures it had taken to address that risk.
Higher costs
Muse Wonder’s API pricing is $4 per million input tokens and $20 per million output tokens, compared with $1.25 and $4.25 for its predecessor. The Decoder reports that Artificial Analysis measured its cost per Intelligence Index task at $4.52, nearly three times Muse Spark 1.3’s $1.60.
Meta plans to use the models to power its Muse agent. The Decoder reports that Muse Wonder will also reach the Muse app, starting with paying subscribers, with specialized models planned for all users over time.
