The work ran from September 19 to 30, 2026, covered 22 experiments and about 160,000 requests to Jev and OpenAI control models, and cost about $18 through early-access API access.
Jev takes text plus structured questions and returns a selected answer, confidence value and probability distribution. The test found accuracy comparable to gpt-4.1-mini when both received class descriptions. The author reports Jev was 2.4 times faster and five times cheaper in that setup. Listed API pricing was $0.042 per million input tokens, with no output charge.
The evaluation also tested confidence thresholds, added classes, long and noisy context, languages, and multiple questions per input. At a confidence threshold above 0.9, 29% of Jev's errors and 83% of gpt-4.1-mini's errors fell above the threshold under the author's measurement method. Forty questions on one text were reported to take nearly the time of one question and cost up to 86% less than separate calls.
The source does not establish universal performance: results are tied to jev-1.13.0, selected datasets, synthetic examples in part, and the author's particular confidence calculation for GPT.