Meta's First Paid AI API Undercuts Every Frontier Lab on Price — and Zuckerberg Announced It Himself

Meta Superintelligence Labs opened the public preview of the Meta Model API on July 9 — the company’s first developer platform where outside engineers can pay to build on a Meta model, after two years of giving Llama away for free. The debut model, Muse Spark 1.1, carries a 1-million-token context window, actively manages that context by remembering earlier actions and compacting history without losing critical steps, and ships built-in web-search grounding that returns cited, real-time answers without a developer building a retrieval stack. Pricing is $1.25 per million input tokens and $4.25 per million output — below every comparable frontier offering that shipped this month. The API is dual-compatible with both the OpenAI SDK and the Anthropic Messages format, so an existing agent points at Meta by changing a base URL and key, not rewriting code. Mark Zuckerberg posted on X to unveil it personally, his first post there in years.

The pricing lands in the middle of a month-long undercutting contest. xAI’s Grok 4.5 shipped six days earlier at $2/$6 per million tokens, with Elon Musk calling it “an Opus-class model, but faster, more token-efficient and lower cost.” OpenAI’s GPT-5.6 launched the same day as Muse Spark across three tiers, from Sol at $5/$30 down to Luna at $1/$6 — meaning Meta’s single model undercuts even OpenAI’s cheapest tier on input while beating xAI’s mid-tier on both ends.

For companies building on model APIs rather than committing to one vendor’s ecosystem, the dual-SDK compatibility is the detail that matters more than the price: switching providers is now a config change, which means the four labs racing on price this month are also racing to make that race easy to win.