EDITION OF OCTOBER 13, 2026
Meta launches Muse Wonder and focuses on AI assistant memory
Meta is launching Muse Wonder and Muse Wonder Full, with access differing by model. The reported results rely on its internal tests. Assistant memory is a central focus of this announcement, with performance still to be independently evaluated.
Legal disclosure: this content was produced by an artificial intelligence system. Editorial transparency
Meta is adding Muse Wonder and Muse Wonder Full to its model lineup. Articles from TechCrunch and The Verge describe distinct access options and a focus on assistant memory. Performance comparisons remain Meta's own, while integration into Muse remains a plan without a detailed timeline.
Meta is expanding its artificial intelligence lineup with Muse Wonder and Muse Wonder Full. TechCrunch and The Verge report the launch of these models in their October 12 articles. Following Muse Spark, described as a lightweight, low-latency model, this new family targets more complex tasks, including programming.
Access differs between the models. Wonder and Wonder Full are available to Muse Code subscribers, according to The Verge. Wonder is also accessible through the Meta Model API. TechCrunch specifies that Wonder Full remains exclusive to subscribers and will not be distributed through this API. This difference matters for developers: integration options vary depending on the model selected.
For Wonder, TechCrunch reports a price of $4 per million input tokens and $20 per million output tokens. These prices describe the unit cost of accessing the model. They are insufficient to determine the cost of a complete task, which would depend in particular on the volumes actually consumed and the results obtained.
Meta presents its models as capable of competing with offerings from OpenAI and Anthropic. The comparisons reported include GPT-6 Astra and Claude Fable 5.1. The documents nevertheless agree on an essential limitation: the cited scores come from Meta's evaluations. They establish what the company claims, without providing independent validation of this equivalence in real-world use.
Memory has a particular role in this announcement. According to TechCrunch, Meta emphasizes the models' ability to retain and synthesize information provided by the user. The company highlights MemoryBench and claims that Wonder outperforms the cited competing models on it. Here again, these are results reported by Meta from its own tests.
This focus suggests that Meta is seeking to make continuity across exchanges a distinguishing feature of its assistants. Better recall of information could facilitate extended tasks, but the available evidence does not allow the reliability of this memory in everyday use to be measured.
Finally, the models are intended to power the Muse agent. The Verge specifies that Meta plans this integration, starting with subscribers, without a detailed timeline. The announced access to the models and their planned integration into the assistant therefore remain separate steps.