Muse Wonder: Meta targets the first row of AI with two models designed for agents

Meta launched Monday, October 12, 2026 Muse Wonder and Muse Wonder Full, two artificial intelligence models designed to perform complex tasks, including through agents. The information of 01net, Next, Courier international and TechCrunch converges on this announcement: after fourteen months of reorganization, the group intends to compete with OpenAI and Anthropic. However, this claimed return is based on evaluations which must be distinguished from the results obtained on a daily basis. [S1] [S2] [S3] [S4]

More powerful models to act in time

Unlike Muse Spark, presented as a light model with a low latency preference, the Wonder family is based on larger models for demanding work. TechCrunch highlights programming in particular; Meta also cites the analysis of legal, financial or health-related documents. The company states that it trained them in its Prometheus data centre, as a follow-up to the redesign of its AI activities. [S4] [S6]

Two capabilities structure this promise. Active Memory should keep important information from conversation to conversation and remind them when it becomes relevant. Active Focus refers to the ability to continue a difficult task, change approach, overcome an error and ask for clarification. Meta also claims that the system asks the user before important actions. Next nuances the novelty of these labels, which he considers marketing: the stake is less the existence of these functions than their comparative effectiveness. [S2] [S5] [S6]

It is also necessary to distinguish the models of Agent Muse, launched before. Wonder and Wonder Full are new engines of the AIA, not a new announcement of this assistant. The October 12 press release predicts that Wonder will soon feed experiences in Muse, starting with subscribers, without giving a detailed timetable for generalization. [S4] [S5]

Results close to leaders, without demonstrated general supremacy

Next reports, based on the results presented by Meta, a score of 56 for Wonder Full in the Artificial Analysis Intelligence Index, compared with 58 for Claude Opus 5.5 and 53 for Claude Fable 5.1 or GPT-6 Astra. Wonder standard reaches 52. These results support a rapprochement with competing models, but not domination over all uses. The first places claimed in certain areas, including the code and the legal ones, remain attributed to the evaluations put forward by Meta. [S1] [S2]

However, the sources do not describe all these comparisons in the same way. International Courier reports an independent test of Artificial Analysis and cites Claude Opus 5.5 Max as the only more powerful model than Wonder Full. Next and TechCrunch place more emphasis on the tests presented by Meta. This difference in attribution and denomination calls for classifications not to be merged without their assessment conditions. Memory and perseverance results also remain measures on specific tests, not a guarantee of reliability for any mission. [S2] [S3] [S4]

Differentiated trade access

Availability is more precise than the expression "two new accessible models". Wonder is offered via Meta Model API and Muse Code; Wonder Full is reserved for Muse Code subscribers, without API access at launch. Next states that model weights are not downloadable and that a local installation is not proposed. This family is therefore distributed in closed form. [S2] [S4] [S5]

Prices differ. 01net refers to a higher cost of use than competition. Next and TechCrunch however give an API rate of 4 dollars per million tokens in entrance and 20 dollars in output for Wonder, which they approximate to that of Claude Opus 5.5. This is not enough to conclude that there is a systematic additional cost. For its part, Courier international highlights the strategic issue of subscriptions: developing revenue from AI beyond advertising and loyalty to Muse users. [S1] [S2] [S3] [S4]

Security, a promise yet to be documented

Meta claims to have trained models to protect private information, refuse harmful demands and resist adverse instructions, while maintaining defensive cybersecurity capabilities. These protections are central to agents who can handle messages, calendars and documents. But the company is still announcing a supplementary safety assessment report. The launch establishes an ambition and the progress claimed; it does not end either the issue of risks or the issue of the lasting reliability of agents. [S5] [S6]