Meta has announced the release of Muse Wonder and Muse Wonder Full, its most sophisticated models to date aimed at automation and coding tasks.
The tech giant said the new models represent a new leap in AI capabilities and will meaningfully improve the performance of its Muse personal AI agents.
Meta said that Muse Wonder scored 78.8 per cent in the coding benchmark DeepSWE 1.1, which tests models on their long-form coding capabilities, while Muse Wonder Full scored 79 per cent.
This puts both models ahead of Gemini 4 Argon’s 77.9 per cent, Claude Opus 5.5’s 74.2 per cent and GPT-6 Astra’s 74.1 per cent.
Meta added that both models can support multi-step automation, with Muse Wonder and Wonder Full having scored 39.1 and 45.4 per cent respectively in the benchmark AutomationBench-AA. This challenges models to complete business tasks in simulated applications such as Gmail, HubSpot, Salesforce, Slack and Zendesk.
In addition to their capabilities on paper, Meta said it has trained both models with “active memory”, to retain user context between sessions over weeks and months and keep agents working on tasks.
It added that another new approach, “active focus”, stop the models quitting on hard tasks and make them more useful for real-world automation.
“Muse Wonder is a frontier-class model and leads on personal and professional agents,” said Alexandr Wang, chief AI officer at Meta.
“Muse Wonder was built to help you do more of what matters, make the hard things easier and bring what you want within reach. This could be doing a cross country move, navigating a tedious insurance filing or turning a hobby you love into a business you own.”
The independent AI analysis firm Artificial Analysis now ranks Muse Wonder Full as the second most capable model for coding, on par with Anthropic’s Claude Opus 5.5.
Meta said both models were trained using its Prometheus data centre, a gigawatt-scale cluster just smaller than Manhattan located in Ohio.
Muse Wonder is available via the Meta API at $4 per million input tokens and $20 per million output tokens, putting it
Muse Wonder Full is only available via Muse Code subscription, at $50 per month in the Power Usage plan and $200 per month in the Maximum Usage plan.
Meta has significantly overhauled its approach to AI over the past year. From 2023 to 2025 it was known for its open, lightweight Llama models that competed with Chinese AI more than Western frontier offerings.
It cancelled this range in September 2025, after acquiring a 49 per cent stake in the AI alignment firm Scale AI for $14.3 billion and onboarding the company’s founder and chief Alexandr Wang as its new chief AI officer. It quickly put Wang in charge of a newly-formed unit, Meta Superintelligence Labs, aimed at producing frontier AI models.
In May, the firm reassigned 7,000 of its employees to work on AI, alongside cuts of around 8,000 workers, according to Reuters reporting. In June, the newswire separately reported that Emily Dalton Smith, the former head of product for Threads at Meta who had been overseeing internal AI restructuring, was leaving the firm.
Mark Zuckerberg, chief executive at Meta, has pledged to produce “personal superintelligence” by producing agents that can automate tasks around the clock.
