Fourteen months after tearing up its AI work and starting again, Meta says it has caught the leaders. On Monday the company unveiled Muse Wonder and Muse Wonder Full, its most capable models yet, claiming state-of-the-art results on personal and professional agent tasks and ahead of OpenAI's and Anthropic's best on several of its own charts.
Mark Zuckerberg told the Financial Times that Meta has caught up with rivals at the frontier of AI, rejecting calls from other industry leaders to slow down. On X he put it simply: "Muse Wonder is a much larger and more advanced class of models that brings us to the frontier".
What the models do
Meta built them for agents: AI that carries out tasks for you rather than only answering questions. According to its research blog, the models were trained at Prometheus, its gigawatt-scale data centre, and rely on two new abilities. Active Memory carries what matters about you across conversations and pulls back the details relevant to the task at hand. Active Focus keeps going when a task gets hard: it tries other approaches, recovers from errors, asks for clarification and checks in before anything important.
Meta also pitches the models at expert work in law, finance and health, such as reading a lease, an insurance claim or someone's finances, and at coding, from small apps to long software projects. Its demos show Muse Wonder planning a browser session, ordering custom merchandise from a person's Instagram photos and building a 3D kitchen game.
Alexandr Wang, Meta's AI chief, told CNBC that the model, codenamed Watermelon, was trained so agents don't stall on everyday jobs: "When you do real-world tasks, you're going to get roadblocks. We specifically tried to train the model to try other approaches instead of stalling and checking back in with the user".
Meta's numbers
The benchmark charts are Meta's own and have not been independently checked. On MemoryBench, an internal test of how well a model remembers a user's context over a long history, Muse Wonder scores 85.6%, against 81.5% for Claude Opus 5.5 and 81.4% for GPT-6 Astra. On AutomationBench-AA, which tests end-to-end business workflows, Muse Wonder Full leads with 45.4%, while the smaller Muse Wonder (39.1%) trails GPT-6 Astra (41.6%) and Claude Opus 5.5 (41.2%).
Meta says the models are its most aligned to date, trained not to share private information without permission, with its strongest safeguards yet against chemical, biological and cyber misuse. A full safety report is promised "soon". That matters for an agent that reads messages, calendars and documents: some Muse users have already complained of the agent acting on their behalf without telling them.
Who can use it and what it costs
Muse Wonder is available now through the Meta Model API and Muse Code, Meta's coding agent. Muse Wonder Full is reserved for Muse Code subscribers, whose plans cost $50 or $200 a month, according to CNBC. On OpenRouter, Muse Wonder is listed at $4 per million input tokens and $20 per million output tokens, with a context window of about one million tokens. Wang told CNBC the API price will still be "meaningfully below" leading models from OpenAI and Anthropic.
For everyone else, Muse Wonder will "soon" power Muse, Meta's personal agent app, starting with subscribers on its $20 or $100 monthly plans. Zuckerberg says Meta will "work to bring this level of intelligence to everyone over time", and plans to use Muse Wonder to build smaller, specialised models for free users. The announcements don't list countries, so availability should follow the Muse app.
No comments yet. Be the first!