Meta has officially unveiled Muse Spark 1.3, the latest flagship model from Meta Superintelligence Labs designed for long-horizon coding and autonomous agentic workflows.

Operating with a 1 million token context window, the model introduces key architecture updates engineered to sustain extended multi-step execution within a single thread. Rather than acting as an overeager automation tool, Muse Spark 1.3 is designed for collaborative reliability: it actively asks clarifying questions when prompts are ambiguous, flags structural uncertainties, and requires user confirmation before taking consequential or irreversible actions. Benchmarks indicate that the model completes complex engineering tasks using roughly 20% fewer tool calls and 25% fewer tokens compared to Muse Spark 1.2. On evaluations like DeepSWE 1.1 (75.4%) and long-context retrieval (98.5%), it posts top-tier performance while competing directly alongside industry peers.

The model specifically targets enterprise software engineers, AI agent developers, and cost-conscious tech teams. By focusing on low tool-call overhead and multi-turn stability, Meta aims to solve the key bottlenecks facing developers building complex refactoring systems, autonomous web agents, and CI/CD tools.

Available immediately via the Meta Model API and integrated into Muse Code, Meta offers standard pricing at $1.25 per million input tokens and $4.25 per million output tokens, along with a discounted $0.10/$0.20 Contributor tier, delivering high-tier reasoning at aggressive operational efficiency.