Meta Releases Muse Wonder, Its First Frontier AI Model

Meta Muse Wonder and a larger sibling, Muse Wonder Full, are the first frontier AI models from Meta. The Meta Superintelligence Lab announced them on Monday, October 12, and TechCrunch reports that Meta describes both as more capable than Muse Spark, the lightweight model it shipped earlier. It is the first frontier launch since Alexandr Wang, the Scale AI founder, restructured Meta's AI work, ProPakistani notes. Meta's own model card places the two models close to Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra on several tests.

How Meta Muse Wonder compares on benchmarks.

On the Artificial Analysis Intelligence Index, Claude Opus 5.5 scores 58, Muse Wonder Full scores 56, Muse Wonder scores 52 and GPT-6 Astra scores about 53, according to ProPakistani. Meta's own figures show Wonder Full leading on long coding and legal agent tasks. It scored 79.0 on DeepSWE 1.1, against about 74 for Opus 5.5 and GPT-6 Astra, and 57.5 on SWE-Marathon, against 46.3 for Opus 5.5. Opus 5.5 still leads several other rows, including HealthBench Pro and Terminal-Bench, so the answer depends on the test a buyer cares about. Most of these numbers come from Meta's testing, and Let's Data Science notes that Meta's full safety report is still pending.

What it costs developers.

Muse Wonder is available through the Meta Model API at $4 per million input tokens and $20 per million output tokens, with cached input at $0.20, according to Let's Data Science. Both models support a context window of about one million tokens and accept text, images, video and audio. Wonder Full is not served through the API at all. It is limited to Muse Code subscribers, whose plans cost $50 or $200 a month. Neither model has downloadable weights. The same outlet estimates the cost per index task at $4.52 for Muse Wonder, against $5.98 for Opus 5.5 and $3.26 for GPT-6 Astra. Wonder Full used 140 million output tokens to complete the index, compared with a median of 88 million, which makes it the more verbose of the group.

The training data question.

The part of the launch that drew the most attention concerns Muse Code. Its subscriptions default to a Contributor mode, under which Meta may use prompts, code, files and outputs to improve its models. Users can switch that off, but the change does not cover content already submitted. Meta's terms reportedly bar sending confidential, sensitive or personal information while Contributor mode is on. The API does not train on customer data, and Contributor mode is not offered for Muse Wonder on the API. Accounts with zero data retention cannot use the mode at all.

What Meta says about the models.

Meta calls the models “the next step on our scaling ladder and a major leap toward personal superintelligence.” The company puts most weight on what it calls active memory, meaning the model keeps and reuses what users share over long periods. Research scientist Kavya Kopparapu said in a launch video that the team wanted the model to remember what matters to users over days, weeks and months. TechCrunch notes that the models were trained at Meta's Prometheus data center in central Ohio.

What is still unknown.

Independent testers have not yet published full results, and Meta's safety report has not appeared. The index rankings are based on one version of the test, so positions could shift as new benchmarks arrive. It is also unclear how many developers will accept the Muse Code training terms. Teams evaluating the model should run a sample of their own prompts through both the API and a rival model, then compare cost per completed task rather than list prices. Developers who check model output as JSON can tidy the responses with the free formatter at /en/json-formatter.