Overview
- Meta released two Muse models on October 12, 2026: muse-wonder-1.0 is available on the Meta Model API at $4 per million input tokens and $20 per million output tokens, while Muse Wonder Full is only included in Muse Code paid tiers.
- Meta’s own 14-row benchmark shows the API-served Wonder wins five rows and loses nine against Anthropic’s Opus 5.5 when matched at the same published price.
- An independent evaluator, Artificial Analysis, placed Wonder lower on its Intelligence Index and measured a much longer time-to-first-token and higher output-token use, which raises latency and per-task cost in agent loops.
- Meta’s model card notes competitor runs may be best-effort and scores vary with quantization, serving settings, prompts, tools and the agent harness, which complicates cross-vendor comparisons.
- Developers are advised to run same‑harness bake-offs on their own repos because subscription gating and differing harnesses mean the best model for real work depends on agent integration, latency and token economics.