Rox says Jev beat GPT-5 Mini on sales reranking by 12%
Three headline numbers from one vendor's internal benchmark, with no dataset or query count attached.
Rox benchmarked two ways of retrieving the context behind its sales agents and reported that Jev classification came out 20x faster, 10x cheaper and 12 percent more accurate than GPT-5 Mini.
We used Jev to retrieve sales data 20x faster, 10x cheaper, and 12% more accurate than GPT-5 Mini.
The setup is retrieval reranking. Rox says every time one of its agents answers a query it pulls from relevant transcripts, emails, CRM notes, news and documents, and it compared LLM-based reranking against Jev classification for picking which of those to surface. Rox describes the aim as serving thousands of points of context so agents can understand accounts, chart relationships and execute on sales.
TypeSafe AI reposted the three figures verbatim, framing them as production-grade reranking for sales data and adding "Are you getting it yet?". That is the vendor amplifying a customer result, not an independent check.
Are you getting it yet?
What is missing
Rox published the headline numbers and nothing underneath them. There is no dataset, no query count, no description of how the LLM reranker was prompted, and no definition of what accuracy means for a ranked list of transcripts and CRM notes. Twelve percent more accurate could be twelve points or twelve percent relative, and the post does not say which. The speed and cost multiples are similarly unanchored, with no latency figure in milliseconds and no price per thousand queries to compare against your own bill.
That matters most for the cost claim. A reranking workload is a high volume one, which is exactly where a 10x gap would show up on an invoice, and also exactly where the shape of the comparison matters. A single frontier-model call reranking a whole candidate set is a different cost profile from one classification call per candidate.
For people building on it, this is one vendor's internal result on its own corpus, not a published eval you can reproduce. The useful part is the shape of the finding, that reranking is a classification-sized job and does not obviously need a general model. The three multiples are Rox's, measured on Rox's data, and nobody outside Rox has seen the setup.