Meta's Muse Spark 1.1 Debuts as Zuckerberg Pitches a 75% Cheaper Frontier Model

Meta unveiled Muse Spark 1.1, the first frontier model from the expensive Superintelligence team Mark Zuckerberg spent the past year assembling. The pitch to developers is blunt: benchmark parity with the top tier of GPT-5 and Claude Fable at roughly a quarter of the API cost. It is the most aggressive pricing move the frontier market has seen this year.

What Muse Spark 1.1 is

Muse Spark 1.1 is a mixture-of-experts model trained on Meta's expanded H100/H200 cluster in Prineville and the new 168MW site in Jamnagar. Meta claims it matches or beats GPT-5-mini on reasoning benchmarks and closes most of the gap to GPT-5 on coding and math, with a 1M-token context window in preview. Weights are not fully open; a research-license variant is available for academic teams.

The pricing pitch

Meta is quoting roughly 75% below comparable OpenAI and Anthropic API prices per million output tokens. That kind of cut usually means one of two things: subsidized launch pricing, or a genuine cost advantage from vertical integration. Meta owns the datacenters, the silicon deals, and the training stack, so it can afford to keep prices low longer than most competitors.

Who this hurts

Cheap frontier tokens are bad news for the mid-tier: OpenAI's mini and nano lines, Google's Flash-Lite tier, and every startup selling a wrapper priced against Anthropic Haiku. Enterprise buyers with high-volume workloads — support automation, document processing, long-context retrieval — should re-run cost models this quarter.

The bottom line

Meta is turning the frontier-model market into a price war it can afford to fight. If Muse Spark 1.1's quality holds up in independent evaluations, expect the entire API price curve to shift down within 90 days.