A little over a year ago, Meta was trailing in the AI race. Its latest models had disappointed. Mark Zuckerberg responded with extraordinary means: he invested more than 14 billion dollars in the company Scale AI, recruited its founder, Alexandr Wang, to head a new laboratory, and poached researchers at great expense.
On Monday, the gamble paid off. Alexandr Wang summed it up in one phrase: Meta's new model rivals the best.
What Meta has released
Meta's laboratory presented two models: Muse Wonder and Muse Wonder Full, the most powerful version. These are its first "frontier" models, meaning capable of rivalling those from OpenAI, Anthropic or Google.
They are designed for agents, those AIs that carry out long tasks on your behalf: handling an insurance claim, understanding a lease, building an application. Meta highlights two new features. An active memory, which keeps track of your exchanges from one conversation to the next. And a focus mode, intended to help the model get out of a dead end by trying another approach.
They understand text, images, sound and video, and can process enormous documents in one go. Their responses, however, remain in text form.
How they fare according to independent tests
On its own tests, Meta places its models at the level of the best, and even ahead of them for long-term memory. As always, these figures need to be checked against an independent evaluation. Here is the general index from Artificial Analysis, which combines ten tests of reasoning, knowledge, programming and professional work.
| Model | Company | General index | Price (input / output, per million tokens) |
|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 58 | $4 / $20 |
| Muse Wonder Full | Meta | 56 | not sold on a usage basis |
| Claude Sonnet 5.5 | Anthropic | 56 | $2 / $10 |
| Gemini 4 Argon | 53 | $2 / $10 (launch price) | |
| GPT-6 Astra | OpenAI | 53 | $10 / $50 |
| Muse Wonder | Meta | 52 | $4 / $20 |
| GPT-6.1 Sol | OpenAI | 52 | $2 / $10 |
The verdict is clear: Meta has indeed joined the leading group. Muse Wonder Full climbs to just behind the leader, and the standard version matches GPT-6.1 Sol. But at 4 dollars for input and 20 for output, Muse Wonder costs twice as much as models of comparable level, such as Sol or Sonnet 5.5.
Meta also claims that Muse Wonder is its safest model, with no dangerous actions during tests in which an agent is pushed to misbehave. A claim that remains to be confirmed by outside evaluators.
How to access it
For the general public, Muse Wonder will soon power Muse, Meta's personal agent, initially for subscribers, at 20 or 100 dollars per month.
For developers, Muse Wonder is available through Meta's API. Muse Wonder Full, meanwhile, is reserved for subscribers to Muse Code, Meta's coding tool, at 50 or 200 dollars a month. It is not sold for general use.
For years, Meta was the great champion of open models. Its Llama models, released from 2023 onwards, could be downloaded and used freely by anyone, and they fed thousands of projects. In July 2024, Mark Zuckerberg even published a letter entitled "Open source AI is the path forward". Muse Wonder breaks with that history: its weights are not downloadable. Meta keeps a small open model, Muse Glimmer, but its best model is now closed, like those of OpenAI or Anthropic.
What this changes
For Meta, it is also a change of business model. The company lives on advertising. With Muse Wonder, it is testing another source of revenue: charging its users, by subscription.
For users, one question remains open: that of data. We recently reminded readers that Meta's apps declare that they collect an enormous amount of data. An agent with active memory, which retains everything from one conversation to the next, makes that question even more important.
What we take away
In fourteen months, Meta has gone from lagging behind to the leading pack. Proof that with enough money and researchers, a company can catch up, even in the AI race.
The club of labs capable of building the best models therefore has one more member, after the release of Gemini 4 Argon this week. But in joining it, Meta has closed the door it had kept open. For those who care about open models, they will now have to look elsewhere, for example towards Mistral and its Large 4.
>>