Muse Spark 1.1: Meta's first paid AI, built for agents

Meta just did something it never has before: put a price tag on its AI. Muse Spark 1.1 launched July 9 in public preview on the new Meta Model API, and it's a pay-as-you-go play aimed squarely at Anthropic and OpenAI's coding business — $1.25 per million input tokens, $4.25 per million output, with $20 in free credits to start.

The model is tuned for agentic work: a 1M-token context window, the ability to delegate to sub-agents running in parallel, and computer use across desktop, mobile, and browser. It zero-shot generalizes to new tools, MCP servers, and custom skills — and knows when to script a task versus click through it. Built under AI chief Alexandr Wang, it's set to replace the Llama models powering WhatsApp, Instagram, and Facebook chatbots.

Meta positions it as competitive with leading frontier models, though its own blog stops short of naming GPT-5.5 or Opus 4.8 head-to-head. A bigger model, code-named Watermelon, is already training.

Why it matters: after telling staff its agent bets were lagging, Meta is now betting its comeback on selling AI — not just giving it away inside apps.