Meta introduces Muse Wonder and Muse Wonder Full, highlighting two new features designed to keep AI agents on task while expanding software and platform capabilities.
Meta unveiled Muse Wonder and Muse Wonder Full, its latest AI models, aiming to reclaim leadership in AI by adding memory and focus controls to agents.
The two standout features are Active Memory and Active Focus, designed to keep AI tasks on track during complex work like coding or legal drafting. Meta says the largest Muse Wonder version performs well on specialized benchmarks and will be available via an API and Muse Code.
Update · Oct 12, 2:00 PM ET
New details from the latest reports show Meta is positioning Muse Wonder as roughly on par with top rivals for coding and professional work, with executives saying parity with ChatGPT and Claude is achievable on those tasks, though no published benchmarks have been released. The stories also flag that Europe may face questions about publishing model weights under the AI Act, and Meta has not yet disclosed whether weights will be released. Separately, pricing details emerged: Muse Wonder will cost $4 per million input tokens and $20 per million output tokens when accessed via the Meta Model API, and Muse Wonder Full will be limited to Muse Code subscribers. Training reportedly took place at Meta’s Prometheus data center in central Ohio.
Meta is pressing its bet on a new generation of AI with Muse Wonder and Muse Wonder Full, the company’s latest push to return to the forefront of artificial intelligence. The rollout arrives after a year that saw a reshaped AI program and a shift toward a tighter safety and reliability framework for powerful agents.
Meta says Muse Wonder will power a range of capabilities across its apps, including a software-writing tool called Muse Code. The company positions the models as part of its broader effort under Meta Superintelligence Labs, a unit that has drawn leadership from the former Scale AI chief to steer its strategy in a crowded AI race.
The standout feature set for Muse Wonder centers on two new ideas: Active Memory and Active Focus. Meta describes Active Memory as a mechanism to help the AI retain relevant context over long, multi-step tasks, while Active Focus is meant to keep the agent tethered to a stated objective and reduce wandering or missteps. In practice, this could help an AI assistant stay on track when it drafts code, analyzes legal documents, or builds software in collaboration with humans.
Meta compares Muse Wonder with the field’s top offerings, saying the largest version of the model performs well on benchmarks that measure capabilities in areas like software creation and handling legal texts. The company also suggests Muse Wonder can stand up to leading models from other labs on several specialized tests, though benchmark results are not always a perfect predictor of real-world results.
Officially, Muse Wonder is being rolled out through an API and via Muse Code, with plans to expand how it integrates into Meta’s own products and services. Meta has not disclosed specifics about training data or model architecture, but framed Muse Wonder as a step toward scaling capabilities safely and responsibly as part of a broader safety framework announced earlier in the year.
Context matters here: Meta has reshuffled its AI leadership and funding in recent years to catch up with rivals like OpenAI,Anthropic, and Google. Muse Wonder is the latest signal that Meta still sees AI as a core strategic pillar, even as the field faces ongoing questions about safety, reliability, and the best ways to deploy powerful agents across products used by millions.
Last month I wrote about how we can build a positive and safe future for everyone: https://t.co/1KtONlqBs5
Every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens.
The…
— Mark Zuckerberg (@finkd) September 15, 2026
For ordinary readers, Muse Wonder could translate into smarter, more capable AI features inside apps you already use. If Muse Code becomes more integrated, you might see AI-assisted coding or software tasks become faster and more reliable. The safety features behind Active Memory and Active Focus are meant to reduce the risk of the AI producing unsafe or off-target outputs as you depend on it for work or personal projects.
Meta says the Muse Wonder family will continue to evolve, with ongoing refinements to safety and alignment. The company emphasized that this is part of a staged approach to bring more capable AI to its platforms while maintaining guardrails and explicit limits on what the models can do.
A set of AI models from Meta designed to power tasks across software, research and apps, with new safeguards and context-management features.
Two features intended to keep the AI on track during long, complex tasks and prevent it from wandering off-target or generating unsafe outputs.
You're reading the quick version.
OpenAI grows in Dublin with 200 new roles and the launch of a second applied AI deployment lab ahead of a major European AI summit.
Google describes how AI aids gene mapping, rapid diagnoses, engineered enzymes, and smarter disease forecasting.
A Santa Monica startup debuts software that converts 2D sketches or PDFs into code-compliant architectural documents, aiming to speed up early design work while keeping licensed professionals in the loop.
The Daily Brief
Every morning, the news that matters from AI, phones, apps and the people shaping tech. Explained in plain English. Free.
One email a day. No spam, unsubscribe anytime. Privacy policy
