Parallel tests eight decision models, with Jev leading early and OpenAI closing the gap after prompt tuning

Parallel, an AI search company, ran a benchmark using real cases from its own search business to compare eight decision models, including Jev, Perplexity and OpenAI Decisions. The evaluation covered four tasks: deciding whether webpage changes matter, filtering low-quality content, choosing a search index and ranking search results. With the original prompts, Jev posted the best average ranking across the four tasks, followed by Perplexity, while OpenAI ranked fifth. The hardest task was deciding whether a webpage change was significant enough to reprocess. In that test, Perplexity scored 93.3, Jev 87.9 and OpenAI 79.0 on a metric that balances false positives and false negatives, where higher is better. Parallel then used Claude to refine prompts for the models. For webpage-change detection, the team prepared 800 manually labeled examples, had Claude review model mistakes and added rules for failure cases over five rounds, keeping only better-performing versions. On an independent test set, OpenAI improved from 79.0 to 88.8, Rune 26B from 74.1 to 83.9 and Nimble 9B from 65.1 to 75.3. Across three tasks combined, OpenAI’s average rose from 88.7 to 95.1, moving it to third after optimization.