DeepSeek Is Now in Claude Code!

So yesterday, DeepSeek released a brand new model,  DeepSeek v3. 1, and we'll be trying out to see how   good it is at coding. They say it's a first step  towards the agent era, which means the model is   going to be really good at tool calling, as many  other agent-focused models are. And we can see   over here that it's scoring better than DeepSeek  v3 and DeepSeek r1 on the software engineering   benchmark, scoring 66, and that is pretty close to  what Sonnet 3. 7 or better than what Sonnet 3. 7 was   scoring. Still not quite there with Sonnet 4 and  Opus 4, but it is better than some of the other   coding models as well. And you can see how much of  an improvement it had over the DeepSeek r1 model,   scoring much better on all the benchmarks,  especially related to agentic tasks. And you can   see the model is much more token efficient than  DeepSeek r1 was, and I'm sure they're going to be   releasing DeepSeek r2 as well in a couple months,  so it'll be interesting to see how they compare   on their benchmarks as well. But of course, the  benchmarks aren't everything, and every model   kind of has its own feel to it, so we'll be trying  it out on a real-world codebase. We're going to be   trying it out via Claude Code, because one new  thing that DeepSeek added is an Anthropic API   endpoint, which means that it's now compatible  with Claude Code. And there's some setup   instructions over here that we'll be going through  shortly. So what we'll be doing is adding to my

application, Tensor AI over here, and it basically  helps you stay up to date with the latest AI   news. You can download it using the link in the  description down below for free. And the thing   that I want to add is like a For You tab over  here, so it knows which articles are relevant to a   user, so they don't have to manually select based  on categories. And I should probably use something   like which articles they engage with previously,  and also their interests as well, and weight it   in some way. And that will require some embeddings  as well, so I'll be using the OpenAI embeddings,   because I'm already using OpenAI in this codebase.   Now, the very first thing we're going to do with   DeepSeek is to have it plan the approach we're  going to take. I have a rough idea of how it will   look like, but I want it planned out in detail.   And I also want to compare the approach that it   suggests to these other models as well. And I'll  be comparing the models via my application called

MindDeck. ai. There's a coupon code down below if  you want to use it to get access. And basically,   it's 100% offline, and you just enter your own  API keys, and everything is stored locally on your   device. It never leaves your device. So we can go  to Settings, and then insert a DeepSeek API key.    So if we go to DeepSeek. platform, and then copy  the key over here, this key will not be working   by the time you're watching it. And then press  Save API keys. Then we can select DeepSeek chat   as one model, DeepSeek Reasoner as another model.   And I also want to use DeepSeek v3 and DeepSeek   r1 as well to see how it compares when it comes to  planning. So I'm going to go to Models over here,   and then find both of the models. So here's  DeepSeek v3. We can press Add over here,   and then search for DeepSeek r1, and then press  Add Model over here. So now I have all the models   loaded up, the non-thinking, thinking, and older  models over here. And I'm basically going to say,   hey, so basically I want to implement a for you  feed in my application. The application is an AI   news application, and it has articles with lots  of different AI news. I want to use OpenAI's   embedding models to embed the articles, and then  recommend relevant articles to users. They put in   their interests on the application during signup,  and because their interests may change over time,   like I need a way of accounting for this. What  would you recommend? I'm using Supabase as my   database, and I want to use PGVector as well to  store the embeddings. Maybe I need an API endpoint   for this. Basically think of all the different  approaches, and then recommend the best approach   for me for this particular application. Press  Stop, and then press Enter. And now it will pass   it to all the models in parallel. And you can  see which models are already faster and which   are slower. Now, after having read through all the  responses over here to the same prompt, I actually   prefer DeepSeek Chat, which is the non-thinking  model that they recently released yesterday,   the most. Because it gives a simple approach,  and then building upon the simple approach,   and then combining it with a hybrid approach as  well. And it gives some example code as well. The   Reasoner model, which they released yesterday, the  thinking model, it answered the question kinda,   and then it just gave me a bunch of code, HTML  code as well, which I did not need. And compared   to the other approaches, I think DeepSeek Chat,  which is running DeepSeek v3. 1, was the most   comprehensive. Anyways, if you're an LLM power  user, and you want a privacy-first way of running   many different models in parallel, to compare  outputs like I just did over here, and to be   able to use many different models, like literally  hundreds of them, and you want all the data stored   locally on your device, rather than on the cloud  somewhere, then you can check out MindDeck. And   there's a link in the description down below, and  there's a coupon code for the first 25 users as   well. It is a one-time subscription, so you don't  have to worry about any recurring monthly fees or   anything like that. Anyways, to run DeepSeek via  Claude Code, you want to make sure you have it

installed, and then run this command over here,  and replace the API key with your API key. Press   run, and then run Claude, and then you should  see something at the top that says that it's   overwritten the base URL, and it's also overridden  some of these environment variables over here. So   we can say, hi, who are you? And you can see that  it says that's Claude Code, because that's baked   into the system prompt. But anyways, we can give  the instructions and the plan that we have from   earlier, and then see how it performs in executing  it. So I'll paste in the plan, and then switch   to plan mode, and press enter over here, and see  what it comes up with. Alright, so it seems that

DeepSeek is done over here, and it took around 17  cents. And I also decided to run Claude Code as   well, using Opus in another folder, which is an  exact duplicate. So we can see how the solutions   compare, because of course, Opus will give a much  better solution. But we want to see how close the   DeepSeek got. So one thing that DeepSeek has done  better, is they added the functions and the tables   in separate files in the schema, like I said in my  claude. md file. Whereas it seems that Claude Opus   did not do that, it just made a single migration  file over here. But looking through the schema   itself, I actually prefer what Opus did, because  they used duration seconds over here, to count   how long they spent in the article, as a way of  knowing how interested they are in it. And when it   actually comes to doing the embeddings, for some  reason, DeepSeek just makes an API endpoint over   here. Whereas Claude Opus, it actually integrates  into an existing workflow. So after the articles   have been generated, right at the bottom, it  triggers an embeddings thing over here, to then   create the embeddings. It does have to resolve the  types, but I actually prefer the overall solution   of Claude Opus over here. And DeepSeek also seems  to have added the API endpoints into a wrong   folder. So it should have added it to a mobile  API folder over here, which Claude Opus did.    But when comparing the API endpoints themselves,  Claude Opus kind of went over the top over here,   because it made a get endpoint, which I wouldn't  think would be required for this particular use   case. We would just need a post endpoint over  here. But as for the recommendations endpoint,   I think Claude Opus was more comprehensive,  because it used a fallback feed over here. And   it also added a recommendation reason too, which I  thought was quite good. But actually neither model   has integrated into the frontend. So my overall  conclusion is a new DeepSeek model is definitely

a step up. DeepSeek chat seems to give better  outputs than DeepSeek Reasoner at least. I think   DeepSeek Reasoner is kind of overcomplicating the  problem. So maybe my prompt has to be adjusted for   this particular model. DeepSeek seems to do okay  in Claude Code. It probably will perform better in   much smaller code bases. This is a monorepo with  many different like segments to it. So it probably   doesn't perform as well because the context window  is smaller. And because the Claude Code system   prompt is probably not fine-tuned for DeepSeek  itself. And when it comes to pricing of the model,   it is pretty good value for money. And they  also have a discount price for these like   off-peak hours over here, which I find quite  interesting. So ultimately, I don't think I   will be using DeepSeek for coding much. But I will  use it for planning though, because I thought that   DeepSeek chat's planning capabilities were quite  interesting. And it did come up with a pretty good   plan. Right now, I currently use O3 when it comes  to planning mostly. So I think probably what I   will do is I will have O3 running on the side over  here, make a new window, and then have DeepSeek   Chat running on the side over here. And then  whatever plan or whatever idea that I'm thinking   of, I'll just paste into both and then have it  run. And then use two hopefully very different   perspectives to like come up with a good plan for  this. Because I think that DeepSeek was trained   on a lot of Chinese content on Chinese social  media, and the Chinese internet and books and   so forth. So it probably has a slightly different  way of looking at things, I imagine. Anyways, if   you do want to try MindDeck, then there's a coupon  code down below in the description for it. It will   get more expensive over time as more features are  added, and the development costs increase and so   forth. But the license is for life, so if you buy  it now, then you will save money in the long run.