Meta launches Muse Wonder models built to remember and finish long agent tasks

The story

Meta has released Muse Wonder and Muse Wonder Full, a pair of artificial-intelligence models designed for agents that remember information across conversations and continue working through long, multi-step assignments. The standard Wonder model is available through the Meta Model API and the company's Muse Code tool. Wonder Full, which Meta presents as the more capable system, is restricted to subscribers on Muse Code's higher-usage plans. Meta says Wonder will later power its Muse personal agent, beginning with subscribers, but has not given a rollout date.

The release is consequential because it marks Meta's clearest attempt yet to compete at the top of the commercial model market after rebuilding its AI organisation. Reuters reported that the company launched the models as it seeks to challenge OpenAI and Anthropic more directly, following heavy spending on infrastructure and research talent. Meta says both systems were trained at its gigawatt-scale Prometheus data centre in Ohio, connecting the launch to the capital-intensive computing strategy behind its AI ambitions.

Meta is pitching two capabilities as the centre of the product. Active Memory is intended to carry relevant facts, constraints and preferences across conversations, then retrieve them when they matter. Active Focus is meant to keep an agent moving toward a goal by trying alternatives, working around blockers, adapting to new information and asking before consequential actions. The company says this design supports work such as interpreting leases, researching financial questions, navigating insurance claims and modifying large software repositories.

The models each offer a one-million-token context window, according to Meta's developer documentation. Muse Wonder costs $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens through the API. Wonder Full has no public token price because access is bundled into Muse Code subscriptions. This split makes the smaller model the route for developers building their own products, while keeping the flagship tied to Meta's subscription environment.

Meta calls the family state of the art across tests covering professional agents, memory, computer use and long-horizon coding. Its published suite includes Harvey's Legal Agent Benchmark, MemoryBench, OSWorld, DeepSWE and SWE-Marathon. Some results are described as independent third-party runs, but the overall comparison is assembled and presented by Meta. The Verge noted that the company is relying on internal testing to argue that Wonder rivals leading OpenAI and Anthropic models, and that its position against Google's newest system remains unclear.

That caveat matters more for an agent than for an ordinary chatbot. A high score on a coding or memory benchmark does not prove that a system will preserve the right facts over months, notice when stored information has become obsolete, recover safely from a failed action or stop before an irreversible step. Real-world reliability also depends on tools, permissions, data sources and the interface that lets a person inspect what the agent remembers and intends to do.

Persistent memory creates an equally important privacy question. Meta says the Wonder family was trained to avoid sharing private information without permission and to check in before important actions. Yet an agent that can work across messages, calendars and documents necessarily handles unusually sensitive context. The value of memory and the risk of memory rise together: useful personalisation requires retention, while trustworthy use requires clear controls for consent, correction, deletion, access logs and separation from other business systems.

The distribution strategy also signals a change in Meta's model economics. The company built much of its recent developer reputation around openly released Llama weights. Wonder is instead delivered as a metered API and a paid coding service, while Meta has not announced open weights for either new model. That approach offers tighter control over safeguards and updates, but it also makes customers more dependent on Meta's pricing, availability and policy decisions.

INNOVOX analysis: Muse Wonder should be judged as an agent platform, not only as a model launch. Meta is combining frontier-model claims with memory, computer use, orchestration and subscriptions—the components needed to own an ongoing working relationship rather than sell isolated answers. If the system can reliably retain context and complete long tasks, Meta gains a credible route to recurring AI revenue and a bridge into its enormous consumer ecosystem. If it cannot, persistent memory may amplify mistakes and privacy concerns rather than create durable value.

What to watch now is independent evidence. Developers need reproducible tests of long assignments, tool failures and changing instructions, plus latency and cost measurements at large context sizes. Users need controls that show what has been remembered and where it will be used. Meta also needs to clarify when Wonder reaches Muse, whether the Full model becomes available beyond Muse Code and how data from personal-agent interactions is governed. Those disclosures will determine whether this is a genuine step toward dependable agents or an impressive benchmark launch awaiting proof in everyday work.

INNOVOX analysis

The most important part of Muse Wonder is not Meta's leaderboard claim. It is the product architecture around the model: persistent personal context, tool use and sustained execution sold through an API and subscriptions. That combination turns the AI race from a contest over one-shot answers into a contest over who can be trusted to hold context and take actions over time. It also moves Meta further from the open-weight posture associated with Llama.

What to watch

Watch for independent evaluations on long-horizon tasks, error recovery and computer use; clear controls for viewing, correcting and deleting remembered information; disclosure of how Muse data is separated from advertising systems; API reliability and pricing under million-token workloads; and whether Wonder Full reaches developers outside Muse Code.