Meta's Muse Spark 1.3 Briefly Tops Google on AI Benchmarks, Undercuts DeepSeek on Price
The AMW Read
Meta's fourth Muse Spark release in five months briefly outranks Google's new Gemini model and undercuts DeepSeek pricing while confirming a planned open-weight strategy, an incremental update for an already-tracked frontier player with segment-level competitive significance.
Named counterparties: OpenAI
Meta's Muse Spark 1.3 Briefly Tops Google on AI Benchmarks, Undercuts DeepSeek on Price
Meta released Muse Spark 1.3, its fourth model iteration in five months after the original Muse Spark in April, version 1.1 in July, and 1.2 in August. On the Artificial Analysis intelligence index the model scored 62, trailing only Anthropic's Claude Fable 5.1 (66) and Claude Opus 5 (63), and briefly overtook Google's Gemini 3.8 Flash (59), which had launched just four hours earlier. Meta Chief AI Officer Alexander Wang publicly mocked Google's release online, while CEO Mark Zuckerberg called the model's performance "beyond imagination." Paired with Meta's Muse Code agent, the company said evaluation results approach Claude Code combined with Opus 5 or Fable 5. API pricing held flat versus 1.2 at $1.25 per million input tokens, $0.15 for cached input, and $4.25 per million output tokens — above Gemini 3.8 Flash but below DeepSeek-V4-Pro's peak rate. Meta said an advanced "max-reasoning" mode remains gated pending safety testing, and it plans larger models plus an open-weight Muse Spark release.
The episode captures how compressed frontier-model release cycles have become: Google's benchmark lead lasted hours before Meta's launch displaced it, in a ranking still topped by Anthropic. Independent developer testing complicated the marketing — one comparison on the MineBench 3D-reasoning benchmark put Muse Spark 1.3's score at 1787 Elo for $6.57 in total cost, versus Gemini 3.8 Flash's 1888 Elo for $1.18, suggesting Meta's price advantage doesn't yet translate into matching output quality at scale. AI Market Watch's earlier coverage of this same release flagged that some of Meta's headline benchmark claims stem from the still-unreleased max-reasoning variant, a distinction that matters more than the leaderboard rank itself.
For builders, headline pricing shouldn't be read as a reliable estimate of agentic task cost — actual runs in the field can cost several times published per-token rates once multi-step tool use and self-correction loops are involved. For investors tracking the frontier race, Meta's cadence of four releases in five months signals aggressive resourcing but still-secondary capability versus Anthropic, with its real edge remaining price rather than quality.
