Particle.news

Meta Releases Muse Spark 1.3 With Long‑Context, Agent and Coding Optimizations

The update brings claimed efficiency gains and a new data‑sharing discount for users while the highest‑capability 'max reasoning' mode remains held for extra safety checks.

Overview

  • Meta announced Muse Spark 1.3 on September 2, 2026, and is rolling it into Muse Code and the Meta Model API with the same $1.25/1M input and $4.25/1M output pricing and a 1 million‑token context window.
  • Meta says the model improves long‑context handling and agent-style workflows and targets coding tasks, reporting about 20% fewer tool calls and roughly 25% lower token use on related workloads.
  • The company introduced a steep discount plan for users who agree to share prompts and model outputs, cutting per‑million token costs to small fractions of the standard rate for contributors.
  • Third‑party benchmarks place Muse Spark 1.3 among leading models and briefly put it ahead of a recent Google Gemini variant, but independent developers report fast, low‑cost inference with mixed results on output quality.
  • Meta is holding back the 'max reasoning' mode pending additional safety testing and says 1.3 includes stronger defenses against adversarial inputs, prompt injection, and confirmation checks before irreversible agent actions.