Meta has taken another step in its attempt to regain prominence in the artificial intelligence race. The company has introduced two new models, Muse Wonder and Muse Wonder Full, positioning them as direct competitors to the best offerings currently available from OpenAI and Anthropic.
The announcement comes in a context where the conversation regarding the most powerful models on the market has been dominated almost exclusively by OpenAI and Anthropic, with Google recently joining the fray with Gemini 4 Argon. Meta is now looking to carve out a space in that dispute after investing billions of dollars in talent and infrastructure for its artificial intelligence division.
What the new models offer and who can use them
As of today, Muse Code subscribers have access to both models. Meta has specified that Muse Wonder will also be available through the Meta Model API, the gateway for developers who wish to incorporate it into their own applications.
The company, however, has left several loose ends. It has not explained exactly what differentiates Wonder from Wonder Full, nor has it provided pricing for API usage, or revealed technical data such as the number of parameters or the context window size for either system.
What Meta has disclosed is the ultimate destination for this technology. The company expects Wonder to eventually be integrated into Muse, its artificial intelligence agent, starting with paid users. From there, it plans to develop specialized models that, in its own words, it can extend to everyone who uses Muse. For now, there is no public timeline for any of these steps.
Comparisons with GPT-6 Astra and Claude Opus 5.5
Meta claims that its new models are on par with OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1. These are the benchmarks the company explicitly cites when presenting Wonder and Wonder Full.
The support for this claim comes from internal tests conducted by Meta itself across various industry benchmarks. This means that, for now, there is no independent verification of these results. It is therefore advisable to take the comparison for what it is: a statement from the company, not a fact verified by third parties.
There is also a significant unknown. It is not known how Meta's models perform against Gemini 4 Argon, the Google system whose real capabilities are still being uncovered, as reported by The Verge, which broke the news.
The context: Meta’s offensive in artificial intelligence
Wonder and Wonder Full do not arrive alone. They are the continuation of Muse Spark, the first model to emerge from Superintelligence Labs, Meta's division revamped for this race, and the Muse agent, which has already generated significant public conversation. Added to that catalog is also Muse Image, the company's image generation model.
The Verge places these launches within an aggressive effort by Meta to regain a top-tier position in artificial intelligence after allocating multi-billion dollar sums to talent and infrastructure. The company's underlying goal is to bring useful AI agents to the general public through Muse.
The path has not been easy. As reported by the same outlet, Meta had to rush to fix a security flaw in Muse just weeks before its launch. To that episode is added another: Muse once sent a YouTuber's postal address to a stranger. Despite these stumbles, the company has moved forward and released the code so that anyone can build devices compatible with Muse.
What changes in the balance against OpenAI and Anthropic
Meta's entry into this dispute has a unique feature: its Muse agent already has active users, giving it a head start over other laboratories that launch models without a mass-market product behind them.
That integration between model and product is precisely the company's strategic bet. Controlling both the model and the application that uses it allows for faster iteration, without depending on third-party agreements. Meta relies on this logic for its plan to have Wonder eventually power Muse from within.
Even so, the announcement leaves more questions open than it resolves. Without performance figures verified by third parties, without public pricing, and without a date for a wider rollout, the comparison with GPT-6 Astra or Claude Opus 5.5 depends on other laboratories or researchers repeating the tests and confirming similar results.
What to review before switching tools
Those who already work with OpenAI or Anthropic models are under no obligation to switch immediately. But Meta's launch leaves some points that should be reviewed before making any decisions.
It is advisable to check if the task that needs to be solved matches the benchmarks Meta used in its internal tests. Since those results have not been independently verified, the most reliable approach is to test the model with your own use cases before trusting the official comparison.
It is also necessary to evaluate whether it is worth subscribing to Muse Code, as the new models are only available to its paid users. Access via the Meta Model API, furthermore, is limited for now to Wonder, with Wonder Full not being available through that channel.
For those developing products, it is recommended to review the terms of the Meta Model API in detail before integrating it into any application, given that Meta has not detailed pricing or usage limits.
Finally, it is important to keep in mind Muse's recent history regarding security and privacy, especially if you plan to connect the agent to personal or customer data.
What this launch means for Muse users
For developers and technical teams, the launch provides another option to test, always with the caution that the comparative results come from Meta itself. Contrasting them with the model currently in use, applied to real tasks, is the only reliable way to know if the switch is worth it.
For Muse users who do not pay for the subscription, the wait will be longer. Meta has confirmed that Wonder will arrive first to those who have Muse Code, and that models intended for all users will arrive later, without a concrete date yet. As long as the company does not publish defined timelines, or until third parties can validate its performance figures, it is most prudent to consider Wonder as a proposal still in the testing phase.
Frequently Asked Questions
What are Muse Wonder and Muse Wonder Full?
They are two new artificial intelligence models from Meta, presented as competitors to OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Claude Fable 5.1.
Who can use these new models?
For now, they are available to Muse Code subscribers. Muse Wonder can also be used through the Meta Model API, intended for developers.
Are Meta’s comparisons with OpenAI and Anthropic verified?
No. The results come from internal tests conducted by Meta itself on various benchmarks, without any independent verification of those figures at this time.
How much does it cost to use Muse Wonder via the API?
Meta has not detailed the pricing for using the Meta Model API or the access limits for developers.
When will Muse Wonder reach all Muse users?
Meta has said that, after reaching paid subscribers first, it plans to extend specialized models to all Muse users, but it has not offered any concrete date.
Has Muse had security problems in the past?
Yes. According to information published by The Verge, Meta fixed a security flaw at the last minute before the launch of Muse, and the agent once sent a YouTuber's postal address to a stranger.
