Meta’s Muse Spark 1.1 large language model has driven recent trader sentiment by posting 62% accuracy on Humanity’s Last Exam leaderboards in September 2026 evaluations, placing it within a few points of Anthropic’s leading Claude Fable 5.1 and Opus 5 variants while outperforming several GPT-5 configurations. This rapid climb from earlier Llama releases reflects Meta’s scaled investment in reasoning-focused training and high-effort inference modes on the 2,500-question expert benchmark, which tests graduate-level knowledge across mathematics, sciences, and other domains. Competitive dynamics favor Meta due to lower inference costs and open-weight options that accelerate developer adoption, though Anthropic and OpenAI maintain slight edges on verified closed-book runs. Key upcoming catalysts include potential Muse Spark 1.2 or 1.3 releases and any year-end model updates before the December 31, 2026 resolution, which could push Meta past the 65% or 70% thresholds if scaling laws and post-training refinements continue at current pace.