TechnologyTracking
Meta Muse Wonder claims SOTA, how does actual performance, reliability, cost compare?
Meta claims its Muse Wonder series achieves state-of-the-art performance on personal and professional agent tasks, but how does its actual performance compare with existing open-source or commercial models, particularly in terms of reliability and cost-effectiveness in real-world scenarios?
Watching: 等待更多独立基准测试和用户实测报告，以及Meta官方安全报告和成本数据更新。
Current read
Meta's Muse Wonder series, released Oct 12, 2026, claims leading benchmarks, but third-party rankings are mid-tier; evidence on real-world reliability and cost-effectiveness is mixed and insufficient.
Evidence shows Muse Wonder series launched Oct 12, 2026, with official claims of near or above Claude Opus 5.5 on intelligence index and Deep SWE benchmarks, but third-party rankings (7th in an 84-model ladder, 9th in 219 models) indicate not top-tier overall. Cost reports are mixed: one notes 60% cheaper than some models, while full version is paywalled with high API pricing, and another mentions increased evaluation costs. No independent data on task completion rates, error rates, or per-task costs is provided, nor direct comparison with open-source models like Qwen, GLM, or DeepSeek. Thus, real-world reliability and cost-effectiveness remain uncertain, requiring more independent benchmarks and user reports.
Key developments
- The assessment representation was corrected; evidence and earlier records are retained.
How we read it
Official benchmarks claim leadership, but third-party rankings are mid-tier: Meta claims 79.0% on Deep SWE for Full, above Claude Opus 5.5, yet third-party ladders rank it 7th of 84 and 9th of 219.
Cost-effectiveness evidence is mixed: one report says Muse Wonder is 60% cheaper than some models, but the full version is paywalled with high API pricing, and another mentions increased evaluation costs.
Lack of independent reliability data: no task completion rates, error rates, or per-task costs are provided, nor direct comparison with open-source models like Qwen, GLM, or DeepSeek.
Earlier updates
- — Public tracking started; no independent probability estimate yet.; Starting public context: Meta said the series achieves state-of-the-art performance on personal and professional agent tasks. Muse Wonder is available via the Meta Model API and Muse Code, while Muse Wonder Full is offered to Muse Code subscribers.
- — Meta's Muse Wonder series, released Oct 12, 2026, claims leading benchmarks, but third-party rankings are mid-tier; evidence on real-world reliability and cost-effectiveness is mixed and insufficient.; Evidence shows Muse Wonder series launched Oct 12, 2026, with official claims of near or above Claude Opus 5.5 on intelligence index and Deep SWE benchmarks, but third-party rankings (7th in an 84-model ladder, 9th in 219 models) indicate not top-tier overall. Cost reports are mixed: one notes 60% cheaper than some models, while full version is paywalled with high API pricing, and another mentions increased evaluation costs. No independent data on task completion rates, error rates, or per-task costs is provided, nor direct comparison with open-source models like Qwen, GLM, or DeepSeek. Thus, real-world reliability and cost-effectiveness remain uncertain, requiring more independent benchmarks and user reports.