Praxist beats Claude Code at 1/12 the spend

Sapient Intelligence’s PRAXIST earned 60 medals on the 75-task MLE-bench suite, including 49 gold, versus Claude Code’s 55 medals and 34 gold. The authors’ MLE-bench paper records model spend of $3,054 against $38,370 for the Claude baseline—about one-twelfth as much—while Sapient’s August 28 announcement frames PRAXIST as a beta research system. The result is an evaluation and cost signal, not an independent production benchmark; teams experimenting with Faraday’s reproducibility-first agent architecture should preserve lineage and human review before treating cheaper agent search as an autonomous lab.