Meta releases Muse Wonder Full, which has a score of 56 on an index
A table that Meta does not operate gives the new model 56. This is the same as Claude Sonnet 5.5 and 2 points below Claude Opus 5.5.
On October 12, Meta released 2 models: Muse Wonder and Muse Wonder Full. Meta says the 2 models are for long tasks. The AI makes steps, uses tools and checks its work. On the Artificial Analysis Intelligence Index, a table that Meta does not operate, Wonder Full has a score of 56. Wonder has 52 and the Muse Spark 1.3 model has 48.
How it works
Muse Wonder is a fast model. Developers can use it through the Meta API and in Muse Code. Muse Wonder Full is a strong model. Only Muse Code subscribers can use it. Meta does not sell it through the API.
The numbers
Wonder Full is in the top 3 on 6 of 10 tests in the index. Claude Opus 5.5 has a score of 58 and Claude Sonnet 5.5 has 56. Meta says that on one law agent test, Wonder Full has a full pass rate of 13.1%. Claude Opus 5.5 has 4.2%.
What is not known
Meta did many of the tests. Some results for other models came from different agent tools. Wonder Full used a very high reasoning level. Meta gave no date for the wide release in its Muse agent. The price of Wonder is $4 for each 1 million input tokens and $20 for each 1 million output tokens.
Insight: Meta gives away the base and sells the top model
Meta made Muse Spark Open, a model with open weights, and gave its Muse agent for free. Its top model, Wonder Full, is only in Muse Code plans of about $50 and $200 each month. This reads as a move to get users with free tools and to charge persons who write code. Look for Wonder Full in the API or in the free agent.
Sources
Posted
