Meta's Muse Spark 1.3 nears the frontier
Meta's Muse Spark 1.3 scores 61 to 62 on Artificial Analysis's Intelligence Index, behind only Anthropic's top Claude models and at a much lower cost per task than rivals at the same score.
Meta has released Muse Spark 1.3, its fourth Muse Spark model in five months. Artificial Analysis reports that the model's gains come mostly from agentic tasks and scientific reasoning.
The generally available variant, Muse Spark 1.3 (xhigh), scores 61 on the Artificial Analysis Intelligence Index. That is up from 57 for Muse Spark 1.2 in August and 53 for Muse Spark 1.1 in July. It ties GPT-5.6 Sol (max), Grok 4.6 (high) and Claude Opus 5 (high). It trails Claude Fable 5.1 (max, 66), Claude Opus 5 (max, 63) and Claude Fable 5 (max, 62). A higher-effort variant, Muse Spark 1.3 (max), is in limited preview for Meta's partners and scores 62.
Tau3-Bench Banking is the biggest driver of the improvement. The xhigh variant rose from 35% to 47%, and the max variant reached 52%, the top score of any model on that benchmark. Terminal-Bench 2.1 rose from 80% to 85%, and the GDPval-AA v2 Elo rose from 1,615 to 1,709. CritPt improved from 18% to 26%, and GPQA Diamond from 90% to 94%. AA-LCR fell 4 points to 79%, and AA-Omniscience accuracy dropped slightly because the model declined to answer more often, which also lowered its hallucination rate.
Pricing is unchanged at $1.25/$4.25 per million input/output tokens, with cached input at $0.15. At those rates the xhigh variant costs $0.55 per Index task. That is the lowest of any model scoring 59 or higher, compared with $0.94 for Grok 4.6 (high) and $0.95 for GPT-5.6 Sol (max). It is still more than Muse Spark 1.2's $0.40, because the new model uses about 57% more input tokens per task.
The model has a 1M-token context window and accepts text, image and video. It is available through Meta's API and Muse Code. Meta has not announced pricing for the max variant.