Meta's Muse Wonder Full tops SWE-Marathon at 57.5 vs Opus 5.5's 46.3, but trails on AA index
daniel_mac8 · x · 2026-10-13
Meta launched the Muse Wonder family with some novel ideas:
- Active Memory: carries context about the user across long durations
- Active Focus: keeps working toward the goal, routes around blockers, and checks in before big actions
- Wonder Full scores 57.5 on Meta's long-horizon SWE-Marathon benchmark vs Claude Opus 5.5's 46.3
Still not quite frontier:
- Opus 5.5 leads Artificial Analysis's index 58 to 56
- Opus wins Terminal-Bench v4.0 (59.6 vs 57.1) and ProgramBench (70.5 vs 59.0)
- Fable 5.5 is expected to drop soon and may widen the gap
The attached charts show Wonder Full also leading on DeepSWE (79.0), AutomationBench-AA (45.4), MemoryBench (85.5) and Harvey LAB-AA (13.1), while trailing on HealthBench Pro, OSWorld and ProgramBench. It's framed as a remarkable comeback from the Llama 4 debacle.
More from coding & agent
- LocalLM Lab SDK adds MCP Apps support, rendering tool UIs inline in chat — AdventurousKeys · 2026-10-13
- NoHumans.Space opens beta: a shared knowledge network where agents pool API findings — dark-epiphany · 2026-10-13
- Stagehand 4.2 adds WebMCP hooks, letting browser agents invoke tools on the fly — jeff_weinstein · 2026-10-13
- Grady Booch: Coding Agents Aren't Even Good at Design, Let Alone Architecture — Grady_Booch · 2026-10-13
- Vals AI: Agent Teams Cost 1.8-5.1x More, Only One Score Gain Was Significant — rseroter · 2026-10-13
- Agent failures increasingly stem from product decisions, not model intelligence — hazelcough · 2026-10-13