Meta shipped Muse Spark 1.3 on September 2, the latest and most capable version of its flagship AI model family. The release, which arrived roughly four weeks after Muse Spark 1.2, represents a significant leap in coding and agentic task performance, with Meta positioning it as a cost-effective alternative to the most expensive frontier models.
On the Artificial Analysis benchmark index, Muse Spark 1.3 scored 61 — tying OpenAI's GPT-5.6 Sol — while costing $0.55 per task compared to $0.95 for the OpenAI model. The model also outperformed Anthropic's Opus 5 on several coding benchmarks, according to early independent evaluations.
The model is available now in Muse Code, Meta's terminal-based coding agent for developers, and through the Meta Model API. It features a 1-million-token context window, allowing it to handle long-running agentic workflows — multi-step tasks where an AI agent iteratively writes, tests, and refines code over extended sessions.
"Muse Spark 1.3 is tuned for the agentic builds developers actually ship," Meta's AI team wrote in the announcement. The model delivers higher first-attempt accuracy on complex coding tasks and better tool-use performance, meaning it can more reliably call external APIs, manipulate files, and interact with development environments.
The release arrives amid intensifying competition in the AI coding space. The model's cost efficiency could be particularly significant for startups and mid-size companies that need frontier-level performance but face budget constraints. Meta's pricing undercuts both OpenAI and Anthropic's premium tiers while matching or exceeding their capabilities on the benchmarks developers care about most.
