Meta launches Muse Wonder models scoring 56 on AI index

Meta released two AI models, Muse Wonder and Muse Wonder Full, on 12 October 2026, with the larger model scoring 56 on the Artificial Analysis Intelligence Index v4.3 at maximum reasoning effort. That score places Muse Wonder Full second on the index, behind only Anthropic's Claude Opus 5.5. The standard Muse Wonder scored 52 on the same index.wired.com+1

What the models score on the index

Muse Wonder Full's score of 56 represents an 8-point improvement over Muse Spark 1.3, which scored 48 on the same index. Muse Wonder Full also ranked in the top three on six of the ten evaluations that make up the Intelligence Index, with its strongest showings in agentic tasks and coding benchmarks. On the aggregate score, Claude Opus 5.5 still holds the top spot.cryptobriefing.com

Meta says the largest version of Muse Wonder outperforms Anthropic's second-best model, Claude Fable 5.1, and ranks just below Claude Opus 5.5 on the Artificial Analysis Intelligence Index. Meta also says the model outperforms OpenAI's GPT-6 Astra and both Claude models on some specialised benchmarks, including ones measuring a model's ability to work with legal documents and build software. It remains unclear how the models stack up against Google's Gemini 4 Argon.theverge.com

The models include two features designed to keep AI agents working toward a stated goal. Active Memory carries relevant details across conversations, while Active Focus keeps the model working when a task gets hard, trying other approaches, recovering from errors, and asking for clarification when needed. These features address a known problem where agents given open-ended commands can stray and produce expensive mistakes or compounded errors.

How the two features keep agents on task

Meta says it trained the Muse Wonder family to avoid sharing private information without permission, turn down requests to assist in harmful acts, and strengthen defences against cyber attacks. The models can still be used by developers and security teams to find flaws in their own code. Meta says it trains, evaluates, and releases its most capable models in line with its Meta Superintelligence Scaling Framework, with a full safety report to be available soon.

Muse Wonder is available through the Meta Model API and Muse Code, priced at $4 per million input tokens and $20 per million output tokens. Muse Wonder Full is restricted to subscribers. Meta says it plans to power its Muse personal agent with the models, rolling out first to subscribers, but has not shared a timeline for that rollout.

Meta did not disclose details of how the Muse Wonder models were built or trained. The company also did not disclose whether it has seen agents powered by Muse Wonder misbehave in testing or what it has done to mitigate that issue specifically. Zuckerberg said in a post on X last month that the company had delayed the release of its Muse models to work on safety.

What remains unconfirmed about the release

Meta announced the first model from its new Superintelligence Labs division, Muse Spark, in April of this year, and has also released a small open-weight model called Muse Glimmer. The company says it plans to release larger open-weight models in the future. Meta restructured its AI efforts in 2025, sinking billions into the newly formed lab and hiring former Scale AI chief executive Alexandr Wang to lead it.