Meta's Muse model undercuts DeepSeek by 29% with data sharing, sparking AI price war

歸藏的AI工具箱 · wechat · 2026-08-06

According to a WeChat article, Meta released new models MuseSpark1.2 and MuseCode coding agent with aggressive pricing. With data sharing enabled, Muse's input/cache/output prices are $0.10/$0.002/$0.20 per Mtok, about 29% cheaper than DeepSeek Flash ($0.14/$0.0028/$0.28) and significantly lower than Luna ($0.20/$0.02/$1.20). Without data sharing, Muse jumps to $1.25/$0.15/$4.25, becoming the most expensive. Benchmarks: AA composite scores are 54.1 (Muse), 49.9 (DeepSeek), 51.2 (Luna); Terminal-Bench: 80.1%, 78.7%, 80.9%; GDPval: 1631, 1558, 1578 Elo. Muse leads overall, Luna slightly higher on Terminal-Bench, DeepSeek's edge is price and open weights. The author notes difficulty accessing Meta's models due to website restrictions. Allowing data training gives a 12x discount, making price comparable to DeepSeek-Flash. As DeepSeek prepares to raise prices, Meta's move may ignite an AI price war.

Related event: Meta Launches MuseCode and MuseSpark, Trades Ultra-low Prices for Training Data(5 posts)→

Original post →

More from Companies & People

Companies & People channel →