Meta releases Muse Wonder, claims AI benchmark gains

Mark Zuckerberg posted on X on Monday, October 12, 2026, the day Meta Superintelligence Lab announced Muse Wonder and Muse Wonder Full. He put the claim in one sentence. "Muse Spark was our first family of models, and now Muse Wonder is a much larger and more advanced class of models that brings us to the frontier," he wrote.

A large language model is an artificial-intelligence system trained on vast amounts of text so it can generate, summarize, and analyze language; the frontier label has come to mean the small set of those models that lead on the hardest evaluations. Zuckerberg’s post fixed the new pair against what had come before and made that tier the stake. The models had names. The chief executive had claimed the ground.

Alexandr Wang posted on X the same day. Muse Wonder was "our largest and most capable model yet," he wrote. It "rivals the big [labs]." Wang, the Scale AI founder Meta had brought in, had restructured the company’s AI work and become chief AI officer over Superintelligence Labs, the group that put Wonder and Wonder Full into the world on Monday. In the time after he took the role, Meta shipped Muse Spark and the Muse agent. What it did not ship were the top-tier frontier launches OpenAI and Anthropic were known for. That held until this release. Wonder and Wonder Full became the first frontier models Meta put out since the restructuring. Both were trained at the company’s Prometheus data center in central Ohio.

Muse Wonder became available through the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, with cached input at $0.20 per million. Muse Wonder Full was limited to Muse Code subscribers and would not be served through the API. Both models would power the Muse agent. Each supported a one-million-token context window and accepted text, images, video, and audio. A token is a small unit of text or other data; the context window is the span of material the model can hold and work with in one pass.

Active memory and active focus sat at the design center. Active memory is the model’s ability to retain and reuse what a user shares over time. Active focus is its ability to keep working through blockers instead of stalling. Research scientist Kavya Kopparapu spoke in the video that accompanied the launch. "We put a tremendous amount of effort into making sure your model can remember what's important to you over days, weeks, months and beyond," she said. Meta’s blog post put the pair in one line. "Together, they mark the next step on our scaling ladder and a major leap toward personal superintelligence."

Meta published a model card and a blog post with benchmark comparisons. Artificial Analysis’s Intelligence Index put Claude Opus 5.5 at 58. Muse Wonder Full scored 56. Muse Wonder scored 52. GPT-6 Astra and Claude Fable 5.1 both sat at 53. Claude Sonnet 5.5 also scored 56. Gemini 4 Argon reached 53. Grok 4.6 landed at 44. Muse Spark 1.3, the earlier Muse model, had stood at 48. The index is an independent aggregate that folds ten evaluations of reasoning, knowledge, coding, and professional work into a single number; higher is better. Wonder Full had closed to within two points of the top mark and tied Anthropic’s mid-tier flagship.

Meta’s own 14-benchmark scorecard put Wonder Full ahead on the longest coding runs and on legal-document agent work. On DeepSWE 1.1, a software-engineering test, it scored 79.0 against roughly 74 for both Opus 5.5 and GPT-6 Astra. On SWE-Marathon, which measures extended coding sessions, Wonder Full reached 57.5 while Opus 5.5 scored 46.3. On Harvey LAB-AA, a legal-agent benchmark, Wonder Full hit 13.1. Opus 5.5 scored 4.2. Those were the rows where Meta’s numbers claimed the sharpest edge.

Claude Opus 5.5 still led several rows on Meta’s chart. It topped HealthBench Pro, the computer-use test OSWorld, Terminal-Bench, and ProgramBench, where it scored 70.5 against Wonder Full’s 59.0. Those were the general and tool-use evaluations where the Anthropic flagship held the edge.

Wonder’s gains against its own predecessor ran sharpest through coding-agent work. On Terminal-Bench 4.0 the score rose from 33 to 57 percent. The Coding Agents Index, which measures how well a model handles complex software tasks on its own, put Wonder Full at 68, tying Claude Sonnet 5.5, while Wonder scored 63, up from 54 for the earlier Muse model. Knowledge remained the thinner side. On the AA-Omniscience test, which checks factual recall, Wonder Full reached 30 and Wonder 25, against 46 for Claude Opus 5.5, 43 for GPT-6 Astra, and 42 for Gemini 4 Argon. Wonder Full cut its hallucination rate—the share of invented or false answers—from 33 to 23 percent versus its predecessor. Gemini 4 Argon still posted the lowest rate, at 15 percent. On ProofBench, a private set of mathematical-proof problems, Wonder Full solves 94 of 100, up from 58 for its predecessor.

Neither model shipped with publicly downloadable weights. The weights remained inside Meta’s own systems. Anyone who wanted the open tier reached it only through the commercial interface or the coding product. Anyone who wanted Wonder Full needed a Muse Code subscription at the higher tier. Both models were set to power the Muse agent, a path that opened first for paying subscribers, with specialized versions planned later for a wider set of users. On launch day the public API accepted traffic for Muse Wonder. Wonder Full stayed gated. The models trained at Meta’s Prometheus data center in central Ohio were the ones now reachable.