A 1M-token agentic model, a paid developer API, and computer-use across desktop, browser, and mobile — all announced in one live event.
What Shipped
Meta announced Muse Spark 1.1, a 1-million-token-context agentic model. Meta reported top scores on MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent V2 — evaluation suites specifically built around multi-step, tool-using agent tasks rather than single-turn chat.
The Bigger Move
The release ships alongside Meta's first-ever paid developer API in public preview (US-only at launch, with $20 in free credits). It features computer-use capability spanning desktop, browser, and mobile environments, alongside parallel subagent delegation — allowing the model to split complex tasks across multiple working subagents simultaneously. Details are available on the Meta AI Portal.
Context
capabilities had stalled for roughly four months — making this product launch a significant competitive comeback.
What It Means for You
This signals Meta's strategy to compete directly on agentic capability rather than standalone chat quality. Engineering teams should evaluate the platform against internal multi-step workflows before taking initial benchmark scores as definitive
