Meta Superintelligence Labs dropped Muse Spark 1.3 this week — the fourth release in five months — and the pitch is squarely aimed at long-horizon agentic work, not single-turn generation. It’s live now in Muse Code and the Meta Model API with a 1M-token context window, so developers can build on it in production today at unchanged pricing: $1.25 per million input tokens and $4.25 per million output.
The practical wins are in efficiency and collaboration. Meta engineers measured roughly 20% fewer tool calls and 25% fewer tokens versus Muse Spark 1.2 on long coding tasks. That matters because agentic workloads bill per round trip — fewer calls and leaner outputs translate directly to lower costs per completed task. On the collaboration side, the model asks clarifying questions when prompts are ambiguous, pulls the user in when it stalls, and confirms before taking consequential actions. It also adapts to how you like to work: frequent status updates or silent background execution.
But there’s a catch baked into the launch scorecard. Meta’s headline benchmark numbers — 75.4 on DeepSWE v1.1, edging out Claude Opus 5 at 74.0 and GPT-5.6 Sol at 72.7 — were measured on the model’s max reasoning mode. That mode is still gated behind safety testing. What developers can actually call today is the xhigh tier, which posts lower scores across the board: 57.2 versus 66.9 on OSWorld 2.0, 1,709 versus 1,754 on GDPval-AA v2 Elo. Artificial Analysis puts the shipping xhigh variant at 61 on its Intelligence Index, level with GPT-5.6 Sol and Grok 4.6, but behind Claude Opus 5 at 63.
This is the third time in the Muse Spark line that Meta has led with a configuration it won’t yet ship, and the pattern is worth watching. The long-context retrieval numbers are genuinely strong — 98.5 on MRCR v2 in the 256K–512K range versus GPT-5.6 Sol’s 91.5 — and that’s where agentic systems live or die. But until max mode clears safety review, the version developers are paying for is not the version Meta is benchmarking. Weights remain closed too, though an open release sits on the roadmap. If the efficiency gains hold up in real-world agent pipelines, that 20–25% token reduction might matter more to your budget than any benchmark delta.
