Transcript
00:00:00So last week DeepSeq released V4 Flash, and I kind of brushed it off as it only scored 52 points on
00:00:04the Artificial Analysis Intelligence Index, which is 8 points behind Kimi K3, and it slots between
00:00:09GLM 5.2 and Gemini 3.6 Flash. The part I completely missed though was its cost. To run the entire
00:00:15test suite at Artificial Analysis, it costs $72 with this model. That is 10 times cheaper than
00:00:21models with the same intelligence scores, and 33 times cheaper than Kimi K3. So this is definitely
00:00:26one to check out, and it's a pretty interesting model, so let's get into it.
00:00:34So the model is called DeepSeq V4 Flash 0731, and the TLDR of the model is it's got 300 billion
00:00:39parameters, with 13 billion of those being active, and has a million token context window, and its
00:00:44weights are on Hugging Face already with an MIT license. If we jump straight to the benchmarks,
00:00:48as I mentioned in the intro, it scores a 52 on the Intelligence Index, so it's not a frontier model,
00:00:53but if we look at this in the context of other models around it, like these 21 here, you can see
00:00:58that it slots in nicely with Gemini 3.6 and 3.5 Flash, GPT 5.6 Lunar on max effort, GLM 5.2 on max,
00:01:05Opus 5 on low, and it's 4 points ahead of Kimi K3 on a low effort. If I switch this over to the
00:01:11agentic index, you can see out of those same models, DeepSeq is actually the third best one. So it's
00:01:15definitely not state-of-the-art, but pricing is where it dominates. If we take those same 21 models,
00:01:20but we look at cost per task, you can see that DeepSeq V4 Flash is actually the cheapest out of
00:01:24all of the models, and GPT 5.6 Lunar that costs the same is actually two points behind on the
00:01:29Intelligence Index. If you compare this to other open models like Kimi K3 on low, which again is worse
00:01:34on Intelligence, it's actually 8 times cheaper, and the same goes for GLM 5.2. It's about 18 times
00:01:39cheaper than the Gemini models, and it gets the same score as them, and obviously it's a lot cheaper
00:01:44than Sonic 5. And all of this is despite the fact that it uses the third most output tokens on all
00:01:49of these tasks across all of the models, which is a pretty crazy result. If we look at a graph of the
00:01:53Intelligence versus cost, you can see it's incredibly close to that most attractive quadrant. Models
00:01:58like GLM 5.2, GPT 5.6 Terra, and Muse Spark are in there, but they're actually only one point ahead
00:02:03of DeepSeq V4 Flash, and that one point difference would come at 9 times the price, so it's probably not
00:02:09worth it. You can even see on a benchmark like RKGI that it stands out in a place of its own,
00:02:13because it's much cheaper than any other model for its performance level, and by the way, on this
00:02:17benchmark, it actually beats Kimi K3 on a max effort, and is 20 times cheaper. As for the raw cost of the
00:02:23tokens themselves, from DeepSeq they advertise 14 cents for a million input tokens, and 28 cents for a
00:02:28million output tokens, but if you go to Open Router, other providers have got this cheaper, and the average
00:02:33is around 8 cents for a million input tokens, and 25 cents for a million output tokens. When you compare this
00:02:38to other models in its intelligence level, it is just so much cheaper. But how well does it do in
00:02:42actual use? I took DeepSeq V4 Flash, and then 7 other similar comparing models, tested them out in
00:02:47open code with the exact same prompt, to create me a personal finance management dashboard, create a
00:02:52full stack application that includes a front end and a back end, and I gave them all an empty folder,
00:02:56so they are starting completely from scratch. Now this is the result I got back from DeepSeq V4 Flash,
00:03:01it actually turned out to be the second cheapest, and also the sixth quickest, so it's a little bit
00:03:05slow there, and I'll show you a full chart at the end of how these models compare. What we can see though is
00:03:09we do have a semi-working CRUD app, there's something going wrong with the responsiveness and the sizing
00:03:13of these cards here, but everything else is working, the charts are pretty nice and working, all of this
00:03:18add funds and send money, I've tested that out and that did work as well, and we can also see that the
00:03:23accounts, transactions, and goals pages here are also working too. As for the UI design of this, I would say
00:03:28it's pretty standard for what I expect out of some of these AI models, maybe the frontier ones are using
00:03:33a bit of a different design these days, but this just has that very AI CRUD app feel. I'm pretty happy
00:03:37with the stack it shows as well, we're using React, although it is using React 18 and not 19, it's using
00:03:42recharts for the charts, which is a sensible choice, and it's also using V on the front end here. For the
00:03:47server, we're using Express and Cores, and bonus points here because it did actually use a real
00:03:52database using Node SQLite. Next up, we have the result that we got back from GPT 5.6 Luna on a
00:03:57max effort, and I've got to say, I had to double check that I chose the right model here, as this
00:04:01UI is just way better than the one we got back from Deep Seek, and this actually only cost me 6 cents
00:04:06and took 6 minutes to make. Now one thing worth knowing is that Open Router currently has a
00:04:10discount on GPT 5.6 Luna, so it is 50% off, so this would have cost 12 cents if we weren't under the
00:04:15discount, so only a few cents more than Deep Seek's, and I think this UI is definitely worth that price
00:04:20difference. As for the features, I have tested, and we can add funds and send money, but we can't go to
00:04:25accounts, transactions, or goals. All of that just has to be on this main page here. It hasn't actually
00:04:29implemented those tabs. It does lose a few points, however, on the stack it chose. It did choose React
00:04:3319, V, and also Express for the server. When it comes to the actual database, it just did all of it in
00:04:38memory in this file here, but obviously this is something that you could modify the original prompt
00:04:42with and just say include a real database. So both of those models cost us under 10 cents, but I want to go to
00:04:47the other end of the spectrum now and take a look at what Opus 5 gave us on a low effort. This cost me
00:04:52$9.65, which is over 160 times more expensive than those other models. You can see here it does have a
00:04:59fairly nice UI. It's a pretty standard CRUD app as we've seen. It's more similar to the Deep Seek one
00:05:03than it is to GPT 5.6, and I will say what Opus has done really well is the UI inside some of these
00:05:09features. So when we transfer or send money, I think this is a really well-built UI, and all of these features
00:05:14work as well. So it's got the accounts, transaction, budgets, goals, and investment pages. When it comes
00:05:19to the actual code, it shows a fairly standard React stack, as we can see here, but where it gets
00:05:23negative points is on the database, because again, it's used an Express database, but this one has used
00:05:28an in-memory data store and not a real database. Now I'm going to go a bit quicker through the other
00:05:31five models. This one here is GLM 5.2 on extra high. It was actually the second most expensive,
00:05:37costing me $3.68, and it also took the longest of 47 minutes. I would say it's done a fairly good job,
00:05:43though. This looks like a standard CRUD app that a lot of these models are producing, and it's got
00:05:46all of the features like dashboard, accounts, transactions, and goals, and we can even click
00:05:50on these and everything is working. It did also choose a pretty good stack for the backend. It
00:05:54actually elected to use Next.js. It's the only model to do so, and it used a real database linked
00:05:59with Prisma. Gemini 3.6 flashes next, and it was actually the sixth most expensive, costing me $1.69,
00:06:05and also the fifth quickest at 11 minutes. Now for the job it's done, I wouldn't say it's great.
00:06:09I mean it does have all of the charts and all of the information, and a pretty similar look to some
00:06:14of the other apps. It's chosen a top nav bar up here, and just overall everything looks a bit more
00:06:19clustered than I think some of the other examples we've seen, and it's a bit of a shame. Gemini
00:06:23used to have the UI advantage, but it's really gone over to OpenAI now. This also uses a pretty normal
00:06:27stack, so it chose React and Express, but it is storing its data in a JSON file. After that we then have
00:06:32GPT 5.5 on a medium effort, and this was actually the fifth most expensive at $1.15, but it was the
00:06:37third quickest, taking only seven minutes. Now it's done a bit of a weird job here compared to the rest
00:06:42of them. It's done everything in a single page. UI design I would say is fairly nice. It is that
00:06:47OpenAI UI design that a lot of these apps have that we've already seen, and everything is working.
00:06:52It's just not really a full stack application, it's more of just a single page app. It did use React
00:06:57and Express for its stack though, but it chose an in-memory datastore. This is the last GPT model
00:07:01that I tested, and it's GPT 5.6 sold on a low effort. This is actually the fourth cheapest,
00:07:05costing me only 90 cents, and it was the quickest, only taking 3 minutes and 42 seconds. You can see
00:07:11it's a pretty similar UI design to Luna, and in my opinion this is the best UI that we got out of all
00:07:15of the models that I tested. And all of these features are working, like the add funds here,
00:07:20but it doesn't have the accounts, transactions, budgets, or goal pages. None of those tabs work.
00:07:24It also picked the pretty standard stack of React and Express, but it does have an in-memory
00:07:28datastore. Finally, the last model that I tested was Kimi K3 on a low effort. This is actually the
00:07:32third cheapest, coming in at 58 cents, and the fourth fastest, taking me 8 minutes. I've got to
00:07:37say though, I'm not blown away by the UI design that we have here. Technically it is all functional,
00:07:41just looks a bit similar to the DeepSeq one, I'd say it's on the same level as that, but all of the
00:07:45features are working, and it actually went with a weird approach for its stack. It chose VanillaJS,
00:07:50so it didn't use any framework, but it did use Express, and it also did have JSON file as its database.
00:07:55I may have gone a bit overboard with the number of tests that I did here, but I really wanted to see
00:07:58how these models with similar scores would stack up. If we take a look at the rankings that I got
00:08:02for my cost per run, I think GPT 5.6 Luna was definitely the winner here. We got a really nice
00:08:07UI design for 6 cents. Again, without the discount, that would cost 12 cents, so it's a bit more
00:08:12expensive than DeepSeat, 50% more, but for that UI design difference, it is definitely a winner.
00:08:17Opus 5, I wouldn't use that for any of these tasks. Opus 5 on a low effort costing me $9.65,
00:08:23when we have models that can cost us under a dollar and give us super impressive results
00:08:27like GPT 5.6 Sol. One thing I will say though is the code base for DeepSeat V4 Flash. The actual
00:08:32backend and the code was really good in comparison to something like GPT 5.6 Luna, where it chose a
00:08:37real database, so maybe I'd use an OpenAI model for the UI design and DeepSeat V4 Flash for some of
00:08:43the backend code, and that way I get the best of both worlds. When it comes to how long each of these
00:08:46models actually ran for, DeepSeat was pretty bad here. It took 23 minutes. When you compare that
00:08:51something like 5.6 Sol, which only took 3 minutes and 42 seconds to create that super impressive
00:08:56result, I do think it could do a bit of work on its speed, but again, since it's so cheap, does it
00:09:00really matter? Overall, I think DeepSeat V4 Flash is a pretty competent model, just maybe not great at
00:09:05UI design, and that's pretty much all I wanted to say. I'm certainly enjoying this trend of the
00:09:09Chinese labs coming in and undercutting the US labs on the pricing and also matching their
00:09:13intelligence, as it might start a price war that benefits all of us, and I think we've already seen
00:09:17that from OpenAI. They have some really cheap pricing for some really competent models. What
00:09:21do you think? Let me know in the comments down below if you would use this model, wait there,
00:09:25subscribe, and as always, see you in the next one.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video