Meta's Muse Wonder Full hits 64.6% on Vals Index with far better token efficiency
EdwardSun0909 · x · 2026-10-13
Vibe Code Bench author Rayan Krishnan's take on Meta's new Muse release: MSL models are pushing the frontier, not just Pareto optimality — Muse Wonder Full scores 64.6% on Vals Index v2.1 vs Claude's 67.0% and GPT-6 Astra's 63.1%. The biggest unlock is token efficiency: WF matches Muse Spark 1.3's performance in half the steps on Vibe Code Bench. Notably, Wonder Full is in-app only (no API), hinting at tighter model–first-party-app coupling or misuse gating. MSL also claims serious alignment progress. The launch post credits the agent team at TBD Labs, one year in, targeting agents you can delegate long-running work to.
Related event: Meta's Muse Wonder Models Win Praise in Early User Tests(7 posts)→
More from coding & agent
- LangChain built SmithDB, a new database just for agent traces in LangSmith — hwchase17 · 2026-10-13
- One dev logged 21,000 commits across 123 repos in October, 327 on his sickest day — doodlestein · 2026-10-13
- rea was built on two $20 Codex Pro plans — devs share token-budgeting lessons — henloitsjoyce · 2026-10-13
- GPT-6 Astra builds 100% 3D hotel website via Blender MCP with 10.7M triangles — ravisparikh · 2026-10-13
- TBD Labs Launches Muse Wonder and Muse Wonder Full, Agents Built for Delegating Real Work — EdwardSun0909 · 2026-10-13
- AI agent gets QuickBooks account locked; ex-Intuit PM explains the rules and 5 safer ways to connect — kiwibooks · 2026-10-13