DeepSeek V4.1 Flash v2 Matches Opus 5.5 for 90% LESS?

Author: Universe of AI · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek V4.1 Flash v2 Matches Opus 5.5 for 90% LESS? · Channel: Universe of AI · 10:55 · Watch on YouTube

Topics: AI evaluation and benchmarks, Software engineering with AI, AI content generation, Model inference and optimization, AI models

Summary

The video reviews DeepSeek version 4.1 Flash, presenting it as a low-cost model with strong coding and agentic-workflow performance. The host compares benchmark results and visual outputs with several frontier models, arguing that quality can look similar despite large price differences. Examples include 3D worlds and game-like outputs, while comparisons also show DeepSeek often takes longer to finish tasks. The main caveat is that the model can overthink, but the host sees its low cost and capable results as likely to intensify competition over AI pricing.

Key points

DeepSeek version 4.1 Flash is described as a minor upgrade, but the host says its practical outputs can show a substantial improvement.
The video highlights strong results in coding-focused and agentic-workflow benchmarks, including Deep Software Engineering.
The host says the model lags behind on Terminal Bench 4.0 compared with some frontier models.
A benchmark comparison presents DeepSeek as achieving competitive pass rates at much lower per-task costs.
A voxel-world example reportedly took about 25 minutes and cost around $5 through the API.
In comparisons with Gemini 4 Argon, GPT-6.1 Soul Ultra, and Opus 5.5, the host says the outputs appear broadly similar, though the frontier models can finish faster.
The video argues that DeepSeek’s low pricing could put pressure on OpenAI and Anthropic to reduce prices.
The host identifies overthinking and longer generation times as a drawback, while suggesting the extra reasoning may help produce strong results.

Chapters

Start	Chapter	Summary
00:00	Introducing DeepSeek 4.1 Flash	The host introduces the model and frames its low price and claimed output quality as the main reasons to examine it.
01:01	Benchmarks and cost efficiency	The video reviews benchmark performance, highlighting coding strengths, a weaker Terminal Bench result, and low costs per task.
04:10	Output examples and model comparisons	The host presents a voxel-world example and compares outputs with Gemini, GPT, and Opus models, emphasizing visual similarity and DeepSeek’s lower cost.
08:20	Pricing pressure and limitations	The video discusses DeepSeek’s potential effect on competitor pricing and concludes that its main trade-off is slower generation linked to overthinking.

Quotes

it is about 90 times cheaper than many of the frontier labs — 00:00

this model excels at coding work — 02:04

In most areas, these outputs are a little bit similar. — 06:45

It does take longer, but you are going to save a lot of money — 08:20

This model thinks a lot. — 10:24

Concepts and entities

DeepSeek V4.1 Flash, Opus 5.5, Gemini 4 Argon, GPT-6.1 Soul Ultra, Deep Software Engineering version 1.1, Terminal-Bench 4.0, DeepSeek, Anthropic, OpenAI

Full transcript (original language)

00:00 It has finally happened. We have a new model from the Chinese labs, and this time around, it is DeepSeek version 4.1 flash. And what’s crazy is that this model is about 90 times cheaper than many of the frontier labs, but it is producing outputs that are pretty much on par to them. So, let’s get into it. Hey, before we get into the video, I just want to take a moment to tell you guys that AI is not slowing down anytime soon. And if you guys want to stay on top of everything that is happening in AI, be on top of all the new model

00:30 drops, then make sure you guys are subscribed to the channel. Your support means a lot, and now, let’s get back into the video. So, the DeepSeek team today dropped their new model, version 4.1 flash, and they mentioned that this is actually a minor upgrade from their old flash model. And if you can look at the benchmarks, we kind of see that this is their older version 4.1 flash, and then this is the one that they dropped on October 17th, which was today. And if you’re watching this video, this model is quite new, so you might not even know that this model is out. The DeepSeek

01:01 team doesn’t do like a great job of marketing themselves as much. They kind of post on X and Hugging Face, and that’s pretty much it. But, I’m going to make a video on it so you guys are aware. What you’re going to see is that this model, obviously, is not going to be compared to Opus 5.5, any of those models, at least on the benchmarks. But, when we look at the output, which I’m going to show you, we’re going to see them side by side against these frontier labs. And what’s crazy to see is that the this model is producing similar outputs to them. And as I said, it’s about 90 times cheaper than many of

01:32 these models in most cases. So, if we look at this benchmark chart that we see at the moment, one important thing to note is that this model is now getting close to Opus 5 and 5.6 sole level, and it’s in some areas actually outperforming it. Especially on the automation bench over here, this model is 60.5, 35.4 on Agent Slasher exam, on the Apex bench, 72.1, and the deep software engineering, this model actually excels pretty heavily on this. This model is outperforming Opus 5 77 on

02:04 this. And this is a pretty, you know, obviously, a very coding-focused benchmark. So, this model excels at coding work. This model excels at agentic workflows, things like that. And where this model does not really do a great job compared to the rest is obviously on Terminal Bench 4.0 over here, 42.3 versus 51.8. So, it’s a bit behind, but not too bad. Then, you know, there’s other benchmarks over here like charts, reading charts with tools and things like that. You can see that this model is getting close to Opus 5 level.

02:35 But, the benchmarks don’t do it justice. I think when we actually look at the outputs, which I’ll show you in a bit, we can really understand why this model is, you know, loved by everyone and why this lab in general is just loved by everyone because they’re able to produce models that are a fraction of the cost of the furniture labs. And in many areas, it might make sense to just use their model because you’re going to be saving a lot on cost. Now, as I said, DeepSeek is known as a lab who is able to give you the most for your dollar. And we can see clearly from this test,

03:05 the Deep Software Engineering version 1.1, what we see at the bottom is the average cost per task and on the left, the pass rate. What we can see over here is that for about a cent, we are getting the model, the low level, getting a average pass rate about maybe like 75 or some sort of percentage. And at the same level, we can see cheaper models from the OpenAI team like GPT-6 Luna at the same price point, not performing as well as the DeepSeek version 4.1 flash model. Not only that, if you want like similar

03:37 level of performance, if you go a little bit to the right, you can see 6.1 Soul at extreme high is actually getting a little bit less of a performance than the one that the low model is getting. But, the version 4.1 flash model, you’re getting a higher level than the GPT 6.1 soul at about 20 cents, which is wild to see. And across the board, you can see even Astra, this model is getting similar levels of scores at like $2 or this is probably close to like $4 and something cents. And Fable 5, which is

04:10 another beast of itself, $20 for about a similar level of pass rate where this model is achieving it actually higher than Fable 5 and is getting it at cheaper cost compared to that. 20 versus about like 10 cents. Isn’t that wild to see? And this is why people love this lab. Now, if we just look at some of the outputs that we are seeing from this model, what we are really going to see is that this model is producing outputs that are pretty much on par. In some areas, does it beat those models like Opus 5.5 and like GPT 6 Astra? No, but

04:42 if I give you these outputs side by side, which I will in a bit, you can see that you can’t really tell a big difference. So, this example that we’re seeing is a medieval voxel world and it was tested on the Deep Sea Harness using the API on max reasoning with a pretty challenging prompt this person mentioned. It took about 2 million voxels in a single file and around 25 minutes, and the cost was only about $5 in API uses. And you can see at the bottom, the output is amazing. I did not expect Deep Sea to be able to produce

05:13 this. One area that they actually mentioned in the release that this model has excelled in is rendering 3D games and things like that. And we can see that translating to tests like these. So, that is why a lot of people are testing this model on just, you know, creating one-shot outputs of 3D games or simulations or worlds, things like that. And I’m pretty surprised with this because this looks amazing. Couple of months back, we were only generating voxel gardens with these type of models. Now, we are generating full-on worlds and simulations, things like that, which

05:43 tells you the rate of development with these models has increased drastically in 2026. And if you want to compare how this model has been upgraded, on the left you see the old version 4.1 flash model, the first output that I’m showing you guys. It’s not bad, it’s okay, you know? And then on the right you can see the level of detail has increased drastically and this is the same prompt. So a lot of people, even though that they actually say that this is a minor upgrade in their blog post, in usability and the output side it’s generating, sometimes that’s translating into more than just a minor update,

06:15 which is wild to see. >> Before we continue, we just launched the Universe of AI newsletter. If you want to stay on top of AI news without having to hunt for it, link is in the description. Don’t miss out. >> Now this test is comparing Gemini 4 Argon, Google DeepMind’s newest strongest model against DeepSeek version 4.1, the 1017 one. So Argon actually took about 13 minutes and it cost about $1.20. DeepSeek did take a longer time, which is something that we are seeing across the board, is that DeepSeek took 15

06:45 minutes and cost 4 cents. But if you look at the output, I’m not being pretty biased or anything like that. In most areas, these outputs are a little bit similar. I would say for sure that Gemini 4 Argon has some of the physics or things like that at a stronger lockdown compared to the DeepSeek version 4.1 flash model, but if I’m just looking at this test one by one and I’m just observing it, it’s hard to tell. You know what I mean? Like if I saw these two tests, I would definitely

07:16 think that they’re from the same model, but as I said, this cost $1.20 versus 40 cents, you know? It’s crazy to see. Here’s a model being compared against GPT-6.1 so ultra and you can see at the bottom the outputs pretty on par. Like DeepSeek is getting an output that’s similar to Soul ultra, you know what I mean? And like I wouldn’t really say that they’re in a different league if I’m just looking at this, even though the benchmarks might scream that, you know what I mean? But if we look at it, DeepSeek managed to achieve this at a

07:48 cost of about 44 cents, and 6.1 Soul was about 90 times more expensive than this, which tells you, once again, that the lab has done it again. They have made a more efficient model compared to anything that’s out there, because they are achieving on par results at 90% cheaper cost. Here’s a model against Opus 5.5. Here’s DeepSeek version 4.1 against Opus 5.5, and DeepSeek once again took a longer time, 1 hour, and cost about 23 cents. Opus 5.5 achieved it in 20 minutes and

08:20 cost about $6. So, as I said, that’s a trend that we are noticing with this model. It does take longer, but you are going to save a lot of money using this model. So, if you have time, and if you’re just working on regular day things, and even sometimes when you’re doing something that is very coding intensive, but if you have more time, you might want to use a DeepSeek version 4.1 model, because it’s going to save you a lot of money. However, it’s going to cost you some time. But if you look at these outputs, they are pretty similar. I’m not trying to gas it up,

08:51 but I just like the DeepSeek output a little bit more, just cuz of the eyes and the generation. It feels more natural compared to the Opus 5.5, but that’s just me being picky. Uh overall, they’re pretty much the same to me. Now, what DeepSeek, the lab itself, does really well every single time is that it starts a pricing war. We saw when they dropped their last flash model and their pro model, it put pressure on many of the frontier labs, and that is why we probably saw, you know, Anthropic drop the pricing for Opus 5.5. Now, the DeepSeek version 4.1 flash model is

09:23 getting at Opus 5 level at 2% of its cost. So, that is once again going to put pressure on these labs, especially OpenAI and Anthropic. And if more Chinese labs drop something more capable, which we are expecting more models from them in a few weeks. That is going to put once again more pressure on these labs to reduce their prices. And even Open AI, they started to serve Soul 6.1 at a very, like, you know, cheaper cost compared to before. So, the pricing war has once again started and it’s all

09:53 thanks to Deep Seek team. So, as I sum it up, this model is basically the state of the art on the deep software engineering. It pretty much mogs all the Chinese models as this person says, while being so cheap. And that is not a surprise from the lab. And it has a faster task completion and shorter outputs. But, in some outputs, we are actually seeing it take longer. So, that is not what the benchmarks actually live up to. In actual production, it looks like this model is taking longer. And I’ll tell you why in a bit. It’s because they have not done a

10:24 great job of fixing overthinking. This model thinks a lot. And that’s something that, you know, people have experienced when they’re working with this model. And that is probably why it takes so long for the model to generate outputs because they’re spending a lot of time thinking. Which could be a good thing, as I said, because you’re getting very close outputs to the Frontier Labs. And that will require to plan out and think properly. So, makes sense. So, remember, this is not going to be a super fast model compared to the Frontier Labs, but it will be definitely a strong model and

10:55 a cheaper model. But, that’s it for today’s video. Make sure you guys are subscribed to the channel. [music] Follow our new newsletter as well at universeofai.beehive.com. As well, subscribe to the main channel World of AI and support us on X by following the Universe of AI Z as well. Until then, I’ll see you guys in the next video.