Meta's Muse Spark 1.3: cheaper, closer to top

The release of Muse Spark 1.3 is the latest in Meta's rapid cadence of model updates—its fourth in five months. The series began in April, with 1.1 in July and 1.2 in August. This iteration shows where Meta has closed the gap and where it hasn't. Its scores climb sharply on tests the Intelligence Index weights most heavily: τ³-Bench Banking, Terminal-Bench 2.1, and GDPval-AA v2. On the banking test, the max tier's 52% score is an outright lead, yet on coding (Terminal-Bench 2.1) and real-world professional tasks (GDPval-AA v2), Claude Fable 5.1 still holds the top position. Meta's gains come with trade-offs: the max variant burns 62% more reasoning tokens than xhigh, and two scores actually dropped—AA-LCR fell from 83 to 79 percent, and factual accuracy in AA-Omniscience slipped by up to three points as the model more often declines to answer when it's unsure.

Price is the counterweight. While the max tier costs more than the $0.40 index-task cost of version 1.2, no rival scoring at least 59 points is cheaper than Muse Spark 1.3's $0.55 per index task. This price point likely positions it as an attractive option despite not being the top performer. Larger models and an open-weights release are said to be on the way.