Muse Wonder Full

Muse Wonder Full, released October 12, 2026 alongside Muse Wonder, was the full version of the Muse Wonder series and the most capable model Meta had shipped at the time. Where Muse Wonder went to the Meta Model API, Full launched for Muse Code subscribers only, inside Meta's coding agent rather than as an API endpoint. Meta used it to argue a point about generalisation: plotted against Muse Spark 1.3 across reasoning efforts, it did better on tasks Meta had not targeted in post-training than its in-distribution scores would predict, which Meta credited to a larger and stronger base model rather than to tuning for the benchmarks.

On Meta's launch table, run at max reasoning, it led every model compared on Harvey's Legal Agent Benchmark as run by Artificial Analysis at 13.1% against 8.6% for GPT-6 Astra, on AutomationBench at 45.4%, on DeepSWE 1.1 at 79.0%, on SWE-Marathon at 57.5% with xhigh reasoning, and narrowly on Finance Agent v2 at 59.1% to Claude Fable 5.1's 58.9%. It did not sweep the table. Claude Opus 5.5 scored higher on HealthBench Professional, JobBench by a tenth of a point, OSWorld 2.1 partial completion, and Terminal-Bench 4.0 as run by Artificial Analysis, 59.6% to 57.1%, while GPT-6 Astra took GDP.PDF at 34.2% to 33.2%. On the Artificial Analysis Intelligence Index it landed at 56, two points behind Opus 5.5.