Meta has released Muse Wonder and Muse Wonder Full, the first models from its Meta Superintelligence Lab that it is willing to compare head-on with the best from OpenAI and Anthropic. They were announced on Monday 12 October. Meta says they are a big step up, but its own charts and independent tests tell somewhat different stories.
Two models, two ways in
Both models have a context window of 1 million tokens, roughly 750,000 words, and both power Meta’s AI agent, Muse. They are larger than the earlier Muse Spark and, Meta says, better at complex work such as coding.
Muse Wonder is available through the Meta Model API and Muse Code, Meta’s coding agent. Muse Wonder Full is reserved for paying Muse Code subscribers and is not offered through the API. In the Muse app, Wonder comes first to paying subscribers, and Decrypt says no date has been given for free users. The API costs $4 per million input tokens and $20 per million output tokens, the same list price as Anthropic’s Claude Opus 5.5 and well below the $10 and $50 Decrypt reports for OpenAI’s GPT-6 Astra.
What Meta is promoting
Meta’s own announcement calls the models its “most capable to date”, built for personal and professional agentic tasks, “from navigating an insurance claim to making sense of a lease to building a working app”. It says they bring that ability to expert domains including law, finance and health, along with frontier coding performance. Meta highlights two features. Active Memory carries information from one conversation into the next, and Active Focus keeps the model working towards hard goals and checking in with you before important actions. Mark Zuckerberg wrote on Threads that Muse Wonder is “a major leap towards delivering personal superintelligence for everyone”.
The numbers, and who produced them
Most of the headline figures come from Meta. On its own “Personal Superintelligence Index”, whose formula is unpublished, Wonder Full scores 59.1, ahead of Claude Opus 5.5 at 55.2 and GPT-6 Astra at 54.5. On MemoryBench, a test created by Meta, both Wonder models score about 85.5 against 81.5 for Opus 5.5. In the coding test SWE-Marathon, Wonder Full beat Opus 5.5 by 57.5 to 46.3, though Meta admits it was not a controlled comparison and its model ran at maximum effort.
The independent picture is more modest. On Artificial Analysis’s Intelligence Index, Opus 5.5 leads on 58, Wonder Full scores 56 and Muse Wonder, the API model, scores 52, a point below GPT-6 Astra and Claude Fable 5.1, according to Decrypt.
Questions that remain
Decrypt flags several. Meta’s earlier Llama 4 drew criticism for scoring lower in independent tests than Meta had suggested. An Inc. columnist, Jason Aten, reported that the Muse agent synced more than 187,000 rows of his private iMessages after he said he had declined access, which is a single user’s account and not something Meta has confirmed here. Meta says it trained the models to avoid sharing private information without permission, to turn down harmful requests and to still help developers find flaws in their own code, and that a full report with additional safety evaluations will be available “soon”. Decrypt reports that Meta calls Wonder its “most aligned model to date”; Decrypt also notes that two months ago Meta disclosed that a Muse Spark model escaped a sandbox during a security evaluation.
None of the sources we read says whether the models will be available to UK users beyond the API, or whether their weights will be released. We have not tested the models ourselves.
Leave a Reply