Models
Meta measures itself against the best AI models, but the tests are its own
Meta says Muse Wonder and Muse Wonder Full are its most capable models and perform almost as well as those from Anthropic and OpenAI. For now, only the company's own tests back that up, and the safety report is still missing.
The essentials
2 confirmed facts · 3 according to the source · 3 open questions
- Meta unveiled Muse Wonder and Muse Wonder Full on October 12, describing them as its most capable models to date.1
- According to the sourceMeta says the larger one beats Claude Fable 5.1 and sits just below Claude Opus 5.5 on a general capability index.3
- According to the sourceAccording to TechCrunch, Muse Wonder costs developers $4 per million input tokens and $20 per million output tokens.4
- According to the sourceMuse Wonder Full is only offered to subscribers of Muse Code, Meta's tool for coding with AI, according to TechCrunch.43
- Meta has not yet published the full safety report for these models; it says it will be available soon.1
Why it matters
Meta had fallen behind Anthropic, OpenAI and Google, according to WIRED3. If its numbers hold up, there is once again a fourth competitor at the top end, and that tends to put pressure on prices and release schedules. The caution lies in where the data comes from: every comparison comes from Meta, and none of the sources reports an independent check.
The other reading has to do with trust. A personal agent works with your messages, your calendar and your documents, as Meta itself acknowledges1. The model is already in developers' hands before the full safety report has been published, which suggests that you still have to take the company's word on that part.
Today, almost nothing, unless you code and pay for Muse Code. If you use the Muse agent, the model will reach you later, subscribers first, and it will remember things about you from one conversation to the next.
The details
Meta has designed them for agents, meaning AI programs that carry out tasks for you instead of just answering. As examples, it gives filing an insurance claim, understanding a rental contract or sorting out someone's accounts, as well as coding. It adds two features: Active Memory, which remembers what matters to you from one conversation to the next, and Active Focus, which aims to keep the model from giving up on a hard task and to have it ask before taking an important step1.
For now they are reaching people who code. Muse Wonder has been available since October 12 in Muse Code, Meta's tool for writing software with AI, and in the API, the route developers use to connect their own apps to the model13. According to TechCrunch, the Full version stays in Muse Code only4. Meta says the model will soon come to Muse, its personal agent, starting with subscribers, without giving a date1.
The comparisons come from Meta's own tests. According to TechCrunch, the company places its models at a similar level to its rivals' and says Muse Wonder beats those from Anthropic and OpenAI on MemoryBench, a test that measures how much a model remembers of what you have told it over time4. WIRED adds that Meta also claims an edge on some specialized tests, and notes that these measurements do not always reflect real-world use3.
On safety, Meta says it has trained the models not to share private information without permission and to refuse harmful requests, and it promises a full report with more evaluations1. According to WIRED, the company has not said whether it saw improper behavior in agents running Muse Wonder during its tests. The same outlet reports that Mark Zuckerberg said on X in September that Meta had delayed its Muse models to work on safety3.
What we don't know
- When Muse Wonder will reach the Muse agent and whether people who don't pay will get it: Meta only says it will be soon and will start with subscribers.
- Whether the results hold up when someone outside Meta repeats the tests.
- What the full safety report will say and whether agents running Muse Wonder have ignored instructions in internal tests.
Background
Meta says it began rebuilding its AI work fourteen months ago1. According to WIRED, Zuckerberg invested billions in Meta Superintelligence Labs and hired Alexandr Wang, former chief executive of Scale AI, to lead it; the first model of that period, Muse Spark, arrived in April3. The race for agents goes back further: on October 2 we reported that OpenAI's Dots deliver, but with flaws.
Sources
- 1Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents
- 2Meta has some Wonder Full new AI models
- 3Muse Wonder Puts Meta Back at the AI Frontier
- 4Meta releases its first frontier model, Muse Wonder
Was this story useful? Yes Not really