Meta's Muse Spark 1.3 targets agentic coding, with a pay-with-your-data tier

Meta has released Muse Spark 1.3, an AI model pitched squarely at agentic workflows rather than one-shot chat. Meta’s framing emphasizes higher first-attempt accuracy and dependable tool calling: the model is built to track context across steps, reconcile conflicting inputs, and pause to ask for direction instead of barreling ahead. It carries a 1M-token context window, multimodal input (video, images, and documents), and coding tuning aimed at longer tasks with fewer wasted turns.

The pricing structure is the more revealing detail. Two tiers ship side by side: a standard ‘muse-spark-1.3’ at $1.25 per million input tokens and $4.25 output, and a ‘contributor’ tier at $0.10 input and $0.20 output — roughly an order of magnitude cheaper — in exchange for opting into ‘product improvement.’ In practice that means the steep discount is paid for with your data, a tradeoff worth weighing before routing any sensitive or proprietary content through it.

Access runs through Meta’s own channels — the Muse Code multi-agent terminal tool and a self-serve Meta Model API — plus OpenRouter for teams that want to slot it into existing setups. Note that the performance and accuracy claims here are Meta’s own; the benchmark and ‘competitive with frontier models’ language is marketing and hasn’t been independently verified.