I just finished testing DeepSeek V3 on several coding tasks and I’m genuinely impressed. This model performs at a level comparable to Claude 3.5 Sonnet, but with one key difference – it’s open source and incredibly affordable.
My first test was creating an image generation site with a built-in prompt enhancer. DeepSeek V3 nailed it on the first try without any guidance from me. Next, I challenged it with a complex text adventure game that lets users type custom actions or pick from suggestions, complete with AI-generated images from Pollinations and GPT-4o Mini as the language model. While no AI has ever gotten this right on the first try, DeepSeek V3 succeeded on the third attempt – one of the best implementations I’ve seen.
For my final test, I had it create a snake game in HTML featuring an AI opponent controlled by an A-star algorithm that moves slightly slower than the player. Again, perfect execution on the first attempt.
The most shocking part? The total cost for all this testing was just one penny. DeepSeek V3 isn’t just matching the capabilities of top closed-source models – it’s doing it at a fraction of the cost. Only Gemini 2.0 Flash is cheaper since it’s free, but it doesn’t perform nearly as well.
This validates what many predicted in 2023: open source models are now catching up to their closed-source counterparts. I plan to use DeepSeek V3 extensively for coding projects going forward. I’ll update this post if I discover any notable strengths or limitations in its performance.
For more context on the rapid advancement of AI models, check out my analysis of Strong AI development here: https://adam.holter.com/strong-ai-is-already-here-look-at-the-evidence/
