Meta's Muse Wonder Full scores 56, tops rivals' flagships

Meta's Muse series has moved quickly. Alexandr Wang said the company originally shipped Muse Spark in April, and that Meta has continued to scale its models across pre-training and reinforcement learning, now reaching the frontier. The October 12, 2026 launch of Muse Wonder and Muse Wonder Full shows that roughly six months of scaling work has been enough to put Meta's models alongside the flagships of Anthropic, OpenAI and Google on the Artificial Analysis Intelligence Index.

Meta is also tying the models' capabilities to two named concepts. Active memory is the ability to record important items about a user and actively search for task-relevant details, while active focus is the ability to keep working on difficult tasks by exploring alternatives and adapting to new information. Meta says the Muse Wonder series combines these two to deliver strong coding ability, and that it also tracked safety metrics such as the rate of executing unapproved code and mishandling personal information, where Meta's models show lower rates than those of Anthropic and OpenAI.

Access is being kept to Meta's paid Muse Code subscription for both models, while Muse Wonder is additionally available through an API. The API pricing Meta lists is $4 per 1 million tokens for input, $0.20 for cached input and $20.00 for output. For businesses already paying for rival coding assistants, this launch gives them a new option to test, though which model performs better on any specific workload is not something the benchmarks alone settle.