Meta's AI research arm, the Superintelligence Lab, has released two new high-end models: Muse Wonder and Muse Wonder Full. According to the company, these are its first models explicitly developed for the most advanced performance tier.
How they differ from earlier models
Meta's prior Muse Spark model was designed to be low‑resource and low‑latency. By contrast, the Wonder family is larger and — the company says — more capable: they report higher performance metrics and improved results on complex tasks such as programming.
In Meta's internal tests, the performance of Wonder and Wonder Full matched that of OpenAI's recently released GPT-6 Astra. Meta described the new models as the next step in its scaling process and a significant move toward developing personal superintelligence.
Recall and comparisons
Meta also stated that the Wonder models outperform Fable, Opus and Astra on measures of information recall (memory) in its comparative evaluations.
Availability and pricing
The Wonder model will be accessible through the Meta Model API. Meta announced pricing of 4 USD (about 1,300 HUF) per one million input tokens and 20 USD (about 6,500 HUF) per one million output tokens. Meta said this cost profile roughly corresponds to the GPT-6 Sol or Opus 5.5 level in terms of price/performance.
Muse Wonder Full will not be available via the API; access will be limited to Muse Code subscribers. Both models will also power the Muse agent.
Training and infrastructure
Meta reported that the models were trained at its Prometheus data center in central Ohio. The Prometheus facility began operating earlier this year.
The Wonder and Wonder Full releases are among the first flagship models Meta has issued since bringing on Alexandr Wang, the founder of Scale AI.
Why this matters
The announcement signals Meta's push into large, high-performance language models and its intent to compete with providers such as OpenAI. The models' technical claims, pricing and the restricted availability of Wonder Full will affect how developers, enterprises and subscribers can adopt and integrate these new models.
