i asked my ai agent to go through every session i've ever had with it, look at all the models i've used, and figure out which ones i actually liked using. not benchmark scores, not reddit opinions, actual sentiment data from my own conversations. across 6500 sessions and nearly 5000 messages analyzed, the results were genuinely surprising.
the best model by raw sentiment was deepseek v4 pro at 81 percent positive. i only used it for 40 sessions but nearly every interaction was good. the most used model was mimo v2.5 with over 500 sessions and 57 percent positive. not the highest percentage but its cost was so absurdly low that its value per positive interaction was untouchable. the biggest money pit was qwen 3.7 max which ate nearly 50 dollars in a month for net negative sentiment. i was actively frustrated with it while paying premium prices. gemini flash preview was the silent killer with over 3000 sessions at 46 percent positive and 200 million input tokens burned on a model i was ambivalent about at best.
the most interesting finding was the disconnect between what i thought was my primary model and what the data showed. kimi k2.6 felt like my daily driver but it was sitting at 54 percent positive while costing significantly more than mimo which performed better sentiment wise. the pattern was clear: expensive models don't automatically mean better experiences. the cheapest model in the lineup was actually delivering the most consistently positive interactions.
based on all this i rebuilt my model stack. mimo v2.5 as the daily workhorse for 65 percent of traffic, deepseek v4 pro for complex coding and reasoning tasks, minimax m3 as a new addition since it beats v4 pro on real world coding benchmarks while being six times cheaper, and kimi k3 reserved only for the hardest problems where nothing else works. also immediately killed qwen 3.7 max and demoted kimi k2.6 from primary to occasional use. the projected monthly cost dropped from about 170 dollars to under 10 dollars while probably getting better results. the lesson is that if you're not tracking which ai models actually make you happy versus which ones just feel premium, you're almost certainly overpaying for worse outcomes.