Meta Launches Muse Spark 1.1, an Agentic Multimodal Model with 1-Million-Token Context

My take: Meta just launched Muse Spark 1.1, its new multimodal model built for complex agentic tasks: task delegation across subagents, computer interface control, and intelligent management of a 1-million-token context window. Mark Zuckerberg returned to X after three years to announce it.

The numbers Meta reports are notable: 88.1 on MCP Atlas (scaled tool use) and 54.7 on JobBench (professional tool use). That said, since Meta is both judge and contestant in its own tests, we cannot know with certainty whether those figures are free from bias. Independent results are worth waiting for. On Terminal-Bench 2.1, which does have external references, Muse Spark scores 80.0 against GPT 5.5's 83.4, which is competitive, not a lead.

What is clear: the model costs $1.25 per million input tokens and is available free in "Thinking" mode on Meta AI. If Meta succeeds in integrating it into tools you already use daily, the access cost drops considerably. How much does a model's price weigh against its real-world performance in your workflow?