Meta's Muse Spark 1.3 sharpens agentic coding

This is a meaningful efficiency-focused release that makes Meta more credible in developer agents, though benchmark wins still need independent real-world validation.

• –Artificial Analysis scores Muse Spark 1.3 xhigh at 61, up four points from 1.2 and tied with several frontier competitors. [Artificial Analysis](https://artificialanalysis.ai/articles/muse-spark-1-3/)
• –Meta’s unchanged API pricing—$1.25 input and $4.25 output per million tokens—gives it a strong cost-performance position for agent workloads. [Artificial Analysis](https://artificialanalysis.ai/articles/muse-spark-1-3/)
• –The model’s 1M-token context, multimodal inputs, and API availability make it relevant for repository-scale coding and tool-heavy workflows.
• –Improved prompt-injection resistance and confirmation around irreversible actions are important safeguards for autonomous agents, while max reasoning remains gated behind additional safety testing. [Meta announcement](https://research.meta.ai/blog/introducing-muse-spark-1-3)
• –Meta’s evaluation methodology notes that third-party model comparisons are best-effort, so developers should test their own repositories and harnesses before switching. [Evaluation methodology](https://research.meta.ai/static/muse-spark-1-3-multimodal-evaluation-methodology)