Meta Takes on Claude and GPT with Two Muse Wonder Models

Meta unveiled two new AI models, Muse Wonder and Muse Wonder Full, on October 12, 2026 (20 Mehr 1405). Designed for personal and professional tasks, the models were trained at the Prometheus gigawatt data center in Ohio and are the result of a fundamental overhaul...

According to ICT News; On October 12, 2026 (20 Mehr 1405), Meta unveiled two new artificial intelligence models named Muse Wonder and Muse Wonder Full. These models, designed to perform individual and professional tasks, were trained at Prometheus' gigawatt data center in Ohio and are the result of a fundamental overhaul of Meta's AI infrastructure over the past fourteen months.

Prometheus data center; 1.3 million GPU infrastructure

Meta confirmed that the models were trained at the Prometheus data center in central Ohio. This data center is a 1 gigawatt center that was launched in 2026 and is equipped with about 1.3 million GPUs. This massive infrastructure has provided the necessary processing power to train new generation meta models.

Active Memory capability; Retention of information in multiple conversations

Active Memory allows models to retain key user information across multiple conversations and retrieve details relevant to the current task. This feature allows users to benefit from a continuous and personalized experience without having to repeat information in each conversation.

Active Focus capability; Testing alternative routes

Active Focus allows models to test alternative routes when encountering obstacles. These models can ask the user for clarification or get confirmation before taking important actions. This approach reduces the risk of taking unwanted actions and increases the user's control over the process.

In the Artificial Analysis Intelligence Index benchmark, Muse Wonder Full scores 56 and Muse Wonder scores 52. Meanwhile, Claude Opus 5.5 is on top with 58 points.

In the SWE-Marathon benchmark, the Muse Wonder Full scored 57.5, which is higher than the Claude Opus 5.5 with 46.3 and the GPT-6 Astra with 42.5. This result shows that the new meta model performs better in long-term programming tasks.

In Harvey Legal Benchmark, Muse Wonder Full scores 13.1, which is much higher than GPT-6 Astra with 8.6 and Claude Opus 5.5 with 4.2.

Muse Wonder is now available through the Meta Model API and Muse Code. The cost of using the API of this model is $4 per million input tokens and $20 per million output tokens. Cached input is also offered for $0.20. This pricing is exactly the same as the Claude Opus 5.5.

Muse Wonder Full is only available to Muse Code subscribers with $50 or $200 plans. Both models support a 1 million token text window.

Some Muse users have reported that the AI agent has read their private messages without permission. Alexander Wang, president of Meta Superintelligence Labs, said that some users had given permission at the beginning of the conversation but had forgotten. This suggests that meta should provide more transparency about how models access user data.

Meta has not yet announced if it will release Muse Wonder weights. Alexander Wang emphasized that the company remains committed to "open source," but it's unclear whether that commitment includes Muse Wonder.

The introduction of Muse Wonder and Muse Wonder Full represents a significant strategic shift in Meta's approach to the AI market. Until now, the company mainly focused on open source models like Llama. But now it's gone straight to war with OpenAI and Entropy, offering API-centric models priced similarly to Claude Opus 5.5.

The smart thing about meta strategy is to focus on long-term tasks. Muse Wonder Full's performance in the SWE-Marathon benchmark, outperforming Claude Opus 5.5 and GPT-6 Astra, shows that Meta is looking to attract developers who need models for complex, long-term projects.

Also, Prometheus' 1GW data center with 1.3 million GPUs shows Meta's commitment to investing heavily in AI infrastructure. This investment could allow Meta to compete with OpenAI and Google in terms of computing power.

However, Meta's main challenge is gaining users' trust about privacy. Reports of personal messages being read by the Muse agent could damage the company's reputation. Meta should provide more transparency about how models access user data and give users more control.

All in all, Muse Wonder and Muse Wonder Full offer new options in the market of artificial intelligence models that can intensify the competition in this field. If Meta can deliver on the performance promises of these models and address privacy concerns, it could gain significant market share.