Show a human reviewer an agent's own explanation of why its action was fine, and the reviewer approves blocked actions almost twice as often.
25.0 percent without the explanation. 46.9 percent with it. If you have a human in the loop anywhere, that number should change how you
DeepSeek AI Just Got WAY More Expensive 😳
DeepSeek just gave their fast agent model eyes, so it can look at a screenshot and build from it.
Before this, it was text only. You couldn't show it anything. Now the image goes straight into the reasoning, not through another model that describes it first. That matters.
I tested Gemini 3.7 vs DeepSeek V4 Flash on a 5-phase CTF challenge!
🎯 🏆 Gemini 3.7: Solved all 5 phases and retrieved the flag. ❌ DeepSeek V4 Flash: Got stuck in the multi-step logic. 🚀 Gemini 3.7 wins this round!
You have reached the end of the archive
All of deepseek