We benchmarked Jev on DigitalOcean Serverless Inference against four open-weight alternatives on 661 real agent tool calls, and d3, a 26B open model, sent 8% of them to human review against Jev’s 50%. When a person reviews escalations, that gap pays for a dedicated H200 above about 380 to 1,070 decisions a day; on inference price alone, serverless Jev stays cheaper at most volumes.