Meta released Muse Spark 1.3, marking its fourth model deployment in five months. Available via Muse Code and the Meta Model API, the release includes the “xhigh” tier, while a compute-heavy “max” tier is currently restricted to a limited partner preview pending further safety evaluations. Independent evaluations from Artificial Analysis place the max tier at 62 points and xhigh at 61 points on the Intelligence Index.
Meta is positioning the model aggressively on price, maintaining rates at $1.25 per million input tokens and $4.25 per million output tokens for the xhigh tier. On specialized benchmarks, Muse Spark 1.3 Max achieved a leading score of 52 percent on τ³-Bench Banking for tool-use agents, while scoring 86 percent on Terminal-Bench 2.1 coding tests and 1,754 on GDPval-AA v2.
The model relies heavily on increased compute, using 62 percent more reasoning tokens in its max configuration compared to xhigh. While the release undercuts market rivals on cost, certain capabilities saw slight dips, with factual accuracy declining on select benchmarks due to the model more frequently declining to answer uncertain queries.
Why it matters
Provides founders with high-performing agentic capabilities at lower token costs than primary competitors.
Demonstrates trade-offs between increased compute spend and incremental gains on complex reasoning benchmarks.
Highlights strong tool-use performance in simulated environment tests relevant to enterprise automation.
Source: the-decoder.com
