wanted a straight comparison between the deepseek v4 flash build i run and the new v4.1 flash, specifically at max versus xhigh reasoning effort. turned out the question had a trap in it. xhigh is not a real setting for either model. both only accept max, high, and low, with high as the default. the older flash release and pro accepted xhigh, and the platform still silently accepts the string for the new builds and quietly routes it below max, so any comparison done that way is not testing what you think it is.
so i ran it properly. eighteen live calls with three repeats per cell, two models times the settings that exist, plus the vendor benchmark table and third party numbers. at max effort v4.1 wins basically every agentic eval. terminal bench 2.1 jumps from 82.7 to 90.6, the engineering benchmark from 54.4 to 74.2, codeforces rating from 3289 to 3471. the ones that stay close are the pure knowledge tests like gpqa at 90.9 versus 89.9. the ones that are not close are the ones that actually matter for an agent, terminal control, security, repository work.
third party index numbers need a warning label. the current index version has v4.1 at 40 and the 0731 build at 35, but an older index version flips the order, 49.9 for the old build versus 39.5 for v4.1. same model names, opposite ordering, different index versions. do not mix those numbers or you will argue with yourself in circles. if i care about agentic work and max effort is on the table, v4.1 is the one to run.