@lord.luke5: deepseek v4.1 flash 1017 just dropped and the cost chart is wild on deepswe it scored 77 percent at about 20 cents a task claude opus 5 scored around 74 percent at over 10 dollars a task on the same test it also beat kimi k3 and glm 5.3 on almost every test here this is deepseek's own chart so take it with some salt and claude opus 5 still wins the hardest stuff like humanity's last exam and terminal bench but that's exactly why I don't run everything through one model I use claude, chatgpt and open source models to get work done running claude or gpt 5.6 all day gets expensive fast so the hard thinking goes to claude and the grind work goes to open source open source also gives me fewer limits I can run it on my own setup no rate limits no surprise bill what model are you using the most right now