Meta introduces Muse Spark 1.3, an agentic coding model with a 1M-token window

Meta has published the model page for Muse Spark 1.3, the newest member of its Muse family of AI models. According to the page, the release is trained for agentic workflows and optimised for competitive coding performance, and developers using it can expect higher first-attempt accuracy and more reliable tool calling than before.

Built for long-horizon agentic work

Meta describes Muse Spark 1.3 as offering a "max reasoning" setting for challenging reasoning and agentic tasks, with improved real-world usability. The model tracks context and prior results, works through messy or conflicting inputs and asks for input when it needs it, the page states. For coding, it is tuned for long-horizon workflows, with fewer unnecessary turns and cleaner output, and performs competitively with frontier models across several coding evaluations, in Meta's words.

Benchmarks and multimodal input

Meta's comparison table sets Muse Spark 1.3 (max) against Muse Spark 1.2 (xhigh), OpenAI's GPT 5.6 Sol (max) and Anthropic's Opus 5 (max) across agentic, long-context and coding suites. On GDPVal-AA v2, an evaluation of knowledge work, the page lists 1754 for Muse Spark 1.3, 1615 for the previous version, 1710 for GPT 5.6 Sol and 1824 for Opus 5. Meta also says the model is natively multimodal: it perceives video, images and documents, and its visual reasoning runs through a real execution environment rather than scripted steps.

Pricing, API and tooling

Both variants of the model carry a one-million-token context window. The contributor tier, whose traffic may be used to improve Meta's products, is listed at $0.10 per million input tokens, $0.002 per million cached input tokens and $0.20 per million output tokens. The standard muse-spark-1.3 endpoint, which Meta says is not used to improve its products, is priced at $1.25, $0.15 and $4.25 per million tokens respectively. Access is offered through the Meta Model API, which is compatible with existing OpenAI SDK clients, and through OpenRouter.

Ways to build

The model page also promotes Muse Code, a multi-agent coding tool for the terminal built for Muse Spark and currently in beta, and a set of cookbooks covering multi-agent orchestration, fanning subagents out into isolated worktrees, computer use, building a GitHub agent with OpenCode and adding search grounding so a model can pull live web information.