Does Jev beat a frontier LLM's accuracy on the same task?

On TypeSafe’s own benchmark across four example workflows, security incidents, trace observability, invoice processing, and customer service, Jev averaged 67.8% accuracy at $0.0004 and 0.4 seconds per case, beating Claude Opus 5 answering the same decisions as a single prompt, which scored 64.8% at $0.34 and 70.5 seconds.

Jev isn’t the most accurate point on the chart: Opus 5 and Sol, another model TypeSafe tested, score higher, 73.1% and 74.1%, once restructured from one prompt into TypeSafe’s own multi-step workflow, which beat the single-prompt version of every model tested on accuracy, cost, and time. What Jev wins outright is cost and speed, roughly two orders of magnitude cheaper and faster than any prompted model on the chart, the same cost asymmetry that favors a classifier over an LLM judge for a narrow, closed-answer question. These are TypeSafe’s own numbers, on its own four example tasks, not an independent benchmark.