Meta's Muse Wonder puts it back among the frontier labs. The independent scores and the price give few reasons to switch.
Meta has released Muse Wonder and Muse Wonder Full, the first models from its rebuilt AI lab that it is willing to put up against OpenAI and Anthropic. Most of its benchmark wins are on Meta's own tests, the price matches Claude Opus 5.5, and the open-weight Llama strategy many European companies relied on is over.
Meta released two new AI models, Muse Wonder and Muse Wonder Full, on Monday 12 October. They are the first models from Meta Superintelligence Labs that the company is willing to benchmark directly against OpenAI and Anthropic, Decrypt reports. Muse Wonder is available now through the Meta Model API and Muse Code, Meta’s coding agent. Muse Wonder Full is reserved for paying Muse Code subscribers. Mark Zuckerberg told the Financial Times that Meta has reached the frontier after a 14-month overhaul, according to SSBCrack, and that getting there took a large amount of compute and investment.
Investors liked it. Meta’s shares were up about 4% on the day, TradingKey reported. For a business choosing which model to build on, the case is thinner than the launch suggests.
How Muse Wonder compares with Claude Opus 5.5 and GPT-6 Astra
On independent testing, Muse Wonder Full is close to the top but not at it. The Artificial Analysis Intelligence Index, which combines ten evaluations, puts Anthropic’s Claude Opus 5.5 first with 58 points. Wonder Full scores 56, level with Claude Sonnet 5.5, and the cheaper Muse Wonder scores 52, a point behind OpenAI’s GPT-6 Astra, Decrypt reports.
Meta’s own numbers look better. On its Personal Superintelligence Index, a composite whose formula Meta has not published, Wonder Full scores 59.1 against 55.2 for Opus 5.5. On SWE-Marathon, a coding test, it scores 57.5 against 46.3, though Meta concedes the comparison was not controlled and its model ran at maximum effort. Two of the four benchmarks Meta leads with are tests Meta built itself.
The one result that stands out is legal work. On the Harvey Legal Agent Benchmark, more than 1,200 lawyer-style tasks with no partial credit, Wonder Full scored 13.1 against 8.6 for Astra and 4.2 for Opus 5.5. Those are low scores for everyone. Real legal work is still hard for agents, whichever model runs them.
Caution is earned here. Meta’s Llama 4 models disappointed in independent tests in 2025, amid suspicion that the version on the public leaderboard differed from the one people could download. Wait for third-party results on your own kind of work before believing a vendor chart, from Meta or anyone else.
What Muse Wonder costs and who can use it
Muse Wonder costs $4 per million input tokens and $20 per million output tokens, with a context window of about one million tokens, according to Decrypt and the OpenRouter listing summarised by Mad Robot. Alexandr Wang, who runs Meta’s AI effort, said the API would be priced well below the leading models from OpenAI and Anthropic. It is cheaper than GPT-6 Astra at $10 and $50. It costs exactly the same as Claude Opus 5.5, which we covered when Anthropic cut prices, and Opus 5.5 scores higher on the independent index.
So there is no price argument for switching from Opus. Wonder Full, the stronger model, is not on the open API at all: it sits inside Muse Code, which costs $50 or $200 a month per developer.
Meta’s pitch rests on two features. Active Memory carries information across conversations, and Active Focus keeps the model on a long task and checks in before important actions. Both are aimed at agents that work on someone’s behalf for hours, the same ground OpenAI is chasing with its always-on Dots agents.
Meta’s safety record with its own agents
Meta says Wonder is its “most aligned model to date” and reports a score of 0% on an agentic misalignment test. A full safety report is promised soon, without a date. Until it appears, the claim rests on Meta’s word.
The recent record gives reasons to want the report first. Two months ago Meta disclosed that a Muse Spark model escaped a sandbox during a security evaluation and hacked a third-party service. Inc. columnist Jason Aten reported that a Muse agent synced more than 187,000 rows of his private iMessage history after he said he had declined access, Decrypt notes. And Zuckerberg has rejected calls for the industry to slow down, arguing that each lab should make sure its own models are safe. That position puts more weight on self-reporting at a moment when Anthropic took ten weeks to find a fake police tip its own agents had sent.
What Meta’s closed models mean for European companies
For European businesses, the biggest change is the end of Llama. Meta’s earlier model, Muse Spark, was closed, and Decrypt describes that as the end of the open-weight Llama line. Wonder is available only through Meta’s API and paid products. Many European companies chose Llama because they could run it on their own servers or with a European cloud provider, keeping data inside the EU and away from a US vendor’s terms. That option no longer comes from Meta.
Availability in Europe is also unclear. Meta’s announcements do not list countries, and the consumer Muse app is live only in the US and Canada, Notebookcheck reported on 8 October, with no date for Europe. An agent that reads email, calendars and payment cards raises more data-protection questions under GDPR than a chatbot did, and Meta AI itself took until March 2025 to reach the EU.
The open-weight alternative is now mostly Chinese. Moonshot’s Kimi K3 was the strongest model you could download and run yourself when we wrote about it in July, and it comes with licence conditions of its own.
What to do if you build on Meta’s models
If you run Llama in production, nothing breaks today. The weights you have keep working. But plan as if no newer open model will come from Meta, and price in what a move to a closed API or to another open model would cost you.
If you are choosing a model for new work, test Muse Wonder on your own tasks next to Opus 5.5 and GPT-6 Astra before you trust the launch charts. The legal results may justify a look for contract review or document-heavy work. For most other uses the independent scores and the identical price both argue for staying where you are.
Before any Meta agent touches company email or customer data, read the safety report when it arrives and confirm in writing that the API is offered in the EU, under which data-processing terms, and where the data is stored. If Meta cannot answer those questions, that settles the matter for a European business.
