Meta Unveils Muse Wonder Model, On Par with OpenAI and Anthropic Models [Watch]

The Muse Wonder and Muse Wonder Full models outperform GPT-6 Astra and Claude Opus 5.5 in some programming tests.

Muse Wonder and Muse Wonder Full models outperform GPT-6 Astra and Claude Opus 5.5 in some programming tests.

Become a special member of Arjan Ambassador in order to benefit from and access special facilities and different departments in Arjan Ambassador.

Meta introduced Muse Wonder and Muse Wonder Full artificial intelligence models. These two models, which are Meta's first attempt to offer the leading models (Frontier), perform better in some programming tests and perform complex tasks than models such as GPT-6 Astra and Claude Opus 5.5. Meta has also developed capabilities to remember user information and continue complex tasks in these models.

The new models have been developed by the meta superintelligence laboratory. Unlike the Muse Spark, which was developed with the goal of providing a light and fast model, the Muse Wonder family is larger and Meta designed it to perform better in complex tasks, especially programming.

According to Meta, the new models are trained using advances in pretraining and reinforcement learning.

The release of Muse Wonder is the first release of the leading meta model after the restructuring of the company's artificial intelligence activities by Alexander Wang, the founder of Scale AI, in August 2025. After these changes, Meta released the Muse Spark light model and the Muse artificial intelligence agent, but so far has refrained from releasing a model at the level of the most advanced products from companies such as OpenAI and Anthropic.

This model scored 57.5% in the SWE-Marathon test, which evaluates the ability of artificial intelligence models to carry out long-term software projects. This figure is equal to 46.3% for Claude Opus 5.5 and 42.5% for GPT-6 Astra.

Muse Wonder Full also scored 79% in DeepSWE v1.1 test. In the same test, GPT-6 Astra scored 74.1% and Claude Opus 5.5 scored 74.2%.

The new Meta model also scored 13.1 on the Harvey LAB-AA test, which evaluates the performance of artificial intelligence in legal tasks. The score of GPT-6 Astra in this test was 8.6 and the score of Claude Opus was 5.5 equal to 4.2.

However, not all test results are in Meta's favor. Muse Wonder Full scored 56 in the Artificial Analysis Intelligence Index and is below Claude Opus 5.5 with a score of 58. In the same index, this model has performed better than Claude Fable 5.1 and GPT-6 Astra, both of which scored 53 points.

In its new models, Meta has emphasized two features named Active Memory and Active Focus. These features are designed to improve the performance of personal AI assistants.

Active memory allows the model to hold important information provided by the user in various conversations and retrieve relevant details when needed. To assess this ability, Meta has pointed to the MemoryBench test, which measures frameworks for continuous learning and memory construction. According to this company's tests, Muse Wonder has scored higher than Fable, Opus and Astra models in this test.

Active focus capability also helps the model keep track of complex tasks, try other solutions when one method fails, and ask the user for clarification or confirmation when needed. Both features were developed with the aim of improving the performance of the Muse personal AI agent. Meta plans to use new models to power this agent as well.

Both Muse Wonder models support a one-million-token context window and can process text, images, audio, and video; However, their output is limited to text.

The Muse Wonder model is now available through the Meta Model API and the Muse Code service. The cost of using it through API is set at $4 for every one million input tokens and $20 for every one million output tokens. The more advanced Muse Wonder Full model, unlike the regular version, is not provided via API and is only available to Muse Code subscribers.

I am a linguistics student and as much as I am interested in "words", I also enjoy exploring the world of technology and science.