Meta unveils Muse Spark 1.3 AI model for coding

Meta has released Muse Spark 1.3, its latest AI model specifically designed for complex coding tasks and autonomous agentic workflows. The update focuses on resource efficiency, claiming a 25% reduction in token consumption and a 20% decrease in tool calls compared to version 1.2. The model is capable of managing multiple workflows within a single long conversation, asking clarifying questions when faced with ambiguity, and requesting confirmation before performing irreversible actions.

To accelerate model training, Meta has introduced a dual-tier pricing structure. While the standard rate remains $1.25 per million input tokens and $4.25 per million output tokens, a new ‘Contributor’ tier offers discounts of up to 95%. Under this tier, input costs drop to $0.10 and output costs to $0.20 per million tokens, provided users agree to share their prompts and generated outputs for Meta’s future training purposes. This strategy aims to secure high-quality coding data, which is often difficult to source.

Performance benchmarks show Muse Spark 1.3 competing closely with industry leaders. It achieved a score of 75.4 on the DeepSWE benchmark, surpassing Claude Opus 5 and GPT-5.6 Sol. However, independent testing from Artificial Analysis suggests a more nuanced landscape, placing its xhigh variant at a 61 on the Intelligence Index, which ties it with GPT-5.6 Sol but trails Anthropic’s Claude Fable 5.1.