Meta's Muse Spark 1.1 can now use your computer — at the cheapest price on the board

Meta shipped Muse Spark 1.1 this month, and two things make it matter beyond the spec sheet. It can operate a computer — clicking, typing, and navigating across desktop, browser, and mobile through unfamiliar interfaces — and it carries a 1-million-token context window it can actively compact mid-task, so it doesn't lose the thread on long, multi-app work. It is also Meta's first model sold through a public API: the company that built its reputation giving models away is, finally, charging for one.

The benchmarks Meta chose to show are all about agents doing jobs, not chatbots answering questions. On JobBench, which measures professional tool use, it posts 54.7 against Opus 4.8's 48.4 and GPT-5.5's 38.3, and Meta claims state-of-the-art results on MCP Atlas, Finance Agent v2, and Humanity's Last Exam with tools. The price is the other half of the pitch: $1.25 per million input tokens and $4.25 per million output — undercutting every frontier lab it lines itself up against.

Read it as the clearest expression yet of the year's real theme: best fit is beating best score. Meta isn't claiming the smartest model on earth; it's claiming the best price for a model that can actually do agentic, tool-using work, aimed squarely at developers building the next wave of computer-use agents. The caveat is the usual one, and it's load-bearing: these are Meta's own comparisons, and computer-use is precisely the setting where small errors compound across steps. The launch table is a starting gun, not a finish line — the number that will decide this is reliability under independent testing, and that data is still weeks away.