I have tried DeepSeek V4 Pro for a personal side project, it is the cheapest model but it was more expensive than a subscription.

I vibe code for fun during my free time from coding for money. I have tried ChatGPT (OpenAI), Claude (Anthropic) and Antigravity (Google). The thing is, when I vibe code I do a lot of things very fast and because I am vibe coding I like to just throw the biggest model with the most context and let it figure stuff out (I know, I will NEVER get a usable product out of such a brainless methodology). The result of vibe coding hard is that I run out of quotas very fast.


I have been for a couple of weeks now been reading and watching videos about how DeepSeek V4 Pro is up there competing with Sonnet and near Opus and near GPT 5.5 but for a fraction of the price (tyhe flash version costs literally 1% of Opus 4.7 when used directly through the API) so I gave it a shot using openrouter.

I topped up my account with $10 and started vibe coding hard. Over weekend the $10 was gone... I know, I could have saved by using the flash model, optimizing planning and building sessions, breaking down tasks, intercalating between different models, including free ones... But like do I wanna bother with that? No I actually do not because my goal is to have fun by watching the AI code and to see things breaking and working at the same time, that is my idea of fun, if I wanted to optimize something or actually solve problems I would just code myself. Vibe coding to me is like turning on a video gaming and spending time without thinking too much, so nope, I am not optimizing my vibe coding strategy.


The conclusion I got is that DeepSeek V4 Pro is NOT CHEAPER THAN THE COMPETITION. I mean yes it is cheaper at the API level cost, but when it comes to competing with the Antigravity, ChatGPT and Claude it does not compete in terms of price. Those subscriptions are heavily subsidized.

The math is this: for around $20 a month I can code for 2 to 3 days before hitting the weekly limits on those other AI labs and for $10 I also ran out of credits within 2 to 3 days using DeepSeek, seems like a bad deal. If I wanted to code daily for a whole month with DeepSeek it wouldhave costed 20 to 30 a week thus 80 to 120 a month.

With the $20 a month from any of the other labs the weekly quota last me the same 2 to 3 days a week with the same volume of work. Also their $100 a month plans have 5x more quotas. At $20 that means 8 to 12 coding days a month and for the $100 monthly it would be 40 to 60 days of coding a week (btw that is why I don't pay the $100 plans, I would be paying for quotas I wouldn't use fully).

In terms of cost it turns out that DeepSeek is not "up to 100x cheaper" as some youtubers and articles claim. If I used the flash model I would have gotten 2x more usage out of it. In the end it is around the same price as the subscription from other labs, if not more expensive!

There, I said it, the controversial conclusion I have got from this experiment: DEEPSEEK IS MORE EXPENSIVE THAN CLAUDE! DEEPSEEK IS MORE EXPENSIVE THAN CHATGPT! DEEPSEEK IS MORE EXPENSIVE THAN GEMINI!

If this is caught by SEO and ends up on Google someone may read it and think "he is dumb, he is just not using it efficiently" but that is the point! I am not using any other AI efficiently! And when not used efficiently DeepSeek is at best the same as any other AI from bigger labs and at worst it is actually double the price!

So in the end I decided that DeepSeek is not worth it. Also, very importantly... If DeepSeek gets stuck in a problem it feels like money is going down the drain as it keeps thinking and tinkering. With other labs subscriptions when GPT, Claude or Gemini (Antigravity) gets stuck at a problem it is quota going down the drain and when the quota ends I just finish the vibe coding session.

Yes, I know, in the end it is all the same, I am paying for the quota and if I waste the quota getting stuck on some problem I am effectively wasting money, but the problem is that with DeepSeek on open router the upper end of spending is unbounded. It feels like gambling "just one more prompt and I am sure we can solve this issue" and after 2 or 3 "one more prompts" you burned a couple more dollars, with other AIs after a few "one more prompts" I am forced to call it quits for the day (there is the option to be billed for extra quotas, I am not stupid, I am not taking this "deal").

How I feel throwing prompts at the AI hoping it will work this time:

=======================================================================