Why don't you use deepseek V4 or kimi k2.5 or 2.6? They are really good models and you will not hit these token limits.