{ "text": "So, Deepseek version 4 Pro is officially\nout today. Now, you might be confused\nbecause this model was technically out,\nbut it was in general availability, but\nthis is the official release, meaning\nthat they have trained the model a\nlittle bit more and produced a stronger\nversion of the model that is available\ntoday. Now, if I were to summarize\ntoday's release in one simple sentence,\nit is that the price to performance is\nbecoming a very, very important thing\nbecause not only do we have a new model\nfrom the Deepseek team, the SpaceX team\nalso dropped Grock 4.6 six. And both of\nthese models are competing with the best\nFrontier Labs at a fraction of their\ncost. Let's start with Gro 4.6. And one\nthing I'm going to say is that I'm\ngenuinely surprised by the SpaceX team.\nI'm not trying to glaze them or Elon\nMusk or anything. I was just very\ncritical of this lab. In 2025, they\ndropped Grog 4 and other models like\nthat, but I wasn't really, you know,\nmind blown with their performance\nbecause they're pretty they're pretty\nsubpar compared to any of the other labs\nout there. But in 2026, it looks like\nthings have kind of changed a little\nbecause Grock 4.5 was quite competitive\nbased off what it cost and Grock 4.6 is\nactually not too bad. If we take a look\nat the benchmarks, for example, if we\nstart with the artificial analysis\nintelligence index, this model achieves\na 61 and Fable 5 is at 62. Now, what's\nreally important to remember is that\nonce again, this model is quite cheap\ncompared to Fable 5. This is about $2\nper million input tokens and $6 per\nmillion output tokens. In Fable 5 sits\nat $10 per million input tokens and $50\nper million output tokens. And this is\nwhy you start to appreciate this release\na little bit more. It might not beat the\nperformance of the best models is\nmatching them. Even GBT 5.6 so which is\na pretty capable model and this is set\nat max on the artificial analysis index.\nThis model achieves 61 and Grock 4.6 61.\nSo it ties it and it's much cheaper. And\nthen even on all of these other\nbenchmarks, for example, the GDP Val\none, it actually beats Fable 5, which is\nat 1741. And then GPT 5.6, it's at 1728.\nI'm not sure why they didn't choose Opus\n5 as well, but I guess they wanted to\nchoose the quote unquote strongest model\nlineup from each lab. And they chose\nFable 5 for Enthropic, which is fair.\nAnd then Deep Software Engineering one,\nwhich is a critical benchmark. This\nmodel doesn't beat Fable 5 or GPT 5.6\nsix soul, but it gets close to it. It's\n65.9 and Fable 5 sits at 70%. But if\nyou're getting results that are pretty\nclose and the model is five times\ncheaper, I wouldn't be too disappointed\nwith this result. And then same thing\nwith Cursor Bench 3.2, the model\nachieves 69.9%,\nfunny number. And Fable 5 sits at 70.5.\nSo once again, closer. And Grock 4.6\nbeats GPT 5.6 Soul. Same thing with the\nFrontier Code. It gets close to Fable 5,\nbeats GPT 5.6 Soul. So yeah, this model\nis actually available in cursor. So the\npartnership with cursor or I guess the\nacquisition has really helped SpaceX\nmake some strides in the AI space this\nyear and Grock build is something that\nyou know maybe not a lot of us have been\nusing so far but it's probably going to\nbe another platform like codeex or cloud\ncode that we start to use but obviously\ncursor is quite strong as well. So you\nhave options available for you to use\nthem in both. And one thing to note is\nthat they're offering two times usage\ninside Grock build and cursor for the\nfirst week. So if you just want to try\nit out, see what you feel about it, then\nyou know it might be worth trying it out\nright now cuz you get double the usage\nin the first week. Now if we were to\ntake a look at some of the outputs that\npeople have been generating with Grock\n4.6, what we're looking at right now is\na Falcon 9 booster return sequence\nsimulation. And this was done in a\nsingle HTML file. And as I said, if you\nexpected to get this type of output from\nGrock in 2025, you would be kind of\nsurprised because you wouldn't expect\nsomething like this to be generated with\nGrock. But now it looks like we have to\nstart taking the Grock team a little bit\nmore serious because this output is\nquite competitive. And obviously, we're\njust looking at a simulation and we're\njust basing it off of a visual\nrepresentation. But if the model is able\nto produce something like this\nconsistently, then I would expect a lot\nof people to start adopting Grock\nbecause number one, it is cheaper than\nthe other labs at least at the moment\nbecause we don't know if this pricing\nstrategy is going to be sustainable for\nthe SpaceX team in the long run. But at\nleast for now, their models are\ndefinitely cheaper compared to the\nothers. And as I mentioned, this model\nexcelled at the artificial analysis\nindex. This model jumped to probably\nnumber four model. It's tied to GPT 5.6\npretty much similar. So you could say\nnumber three as well, but the models\nbefore that are Fable 5 and Opus 5,\nwhich are only above the model by about\na 1% or a 2% difference. So yeah, even\non this intelligence index, which if\nyou're not familiar with has nine\nevaluations. So on all of these\nevaluations, it's kind of matching\nalmost Fable 5 performance, which is\ncrazy to see. Before we continue, we\njust launched the Universe of AI\nnewsletter. If you want to stay on top\nof AI news without having to hunt for\nit, link is in the description. Don't\nmiss out. And what you see on screen\nright now is a racing game that Grock\n4.6 build. And based off of this post,\nthe model took about 1 minute and it was\na fiveword prompt, which was a create a\nsimple racing game in HTML. So, if\nyou're able to generate something like\nthis easily using Grock 4.6, I think a\nlot of people will be happy. And this is\na more detailed analysis of what it\ncosts to run GPT 5.6 6 on the artificial\nanalysis index and what it produced\nmeaning the output. Both of these models\nif you remember scored 61 on the\nartificial analysis index. Now to run\nthe whole test with GPT 5.6 it cost\nabout 2.8K and then with Grock 4.6 it\ncost about 1.1Kish. And this tells you\nthat you're getting similar level of\nperformance at half the cost. So yeah,\nthis is a big release for the SpaceX\nteam because they just proved once again\nthat they are a lab that you seriously\nstart need to considering, especially in\n2026. And I'm going to talk more about\nDeep Seek version 4 Pro GA. But\nbasically what we're seeing today is\nthat both of these releases kind of\nemphasize the fact that performance and\nall above that is the price at what\nyou're getting for that performance is\nbecoming more and more important for all\nusers because we see many labs now\nfocusing on creating the best model at\nthe cheapest cost. Last year in 2025\nmost of the labs were just focused on I\nwould say creating the strongest model.\nYes, cost was important, but I think\nmost of the times the frontier labs,\nmeaning OpenAI, Enthropic, were kind of\nmore lenient on that fact because they\ndidn't have as strong of a competition.\nIntelligent models that are maybe not\nalways ahead of OpenAI Enthropic, but\nmatch their performance at a fraction of\nthe cost. So yes, price toerformance\nratio is becoming a critical I would say\nindicator in 2026. Now, this is the\nupdated benchmark chart after the\nrelease of the new model. And one thing\nyou'll see across the board is that it\nmatches the top level performance of\nmany of the models. The one thing\ninteresting over here is that they\nhaven't put Opus 5 here for some reason.\nThere is Fable 5 here that we can\ncompare this model against. But one\nthing you'll notice is that Deep Seek,\nremember this model costs 43.5 cents per\nmillion input tokens and 87 cents per\nmillion output tokens. While the other\nmodels all over here are way more\nexpensive than that. So the first thing\nif you look at terminal bench 2.1 the\nmodel scores 87.9.\nThe older version of the model was 72.1\nand the flash version which we got last\nweek was 82.7. And what's crazy is that\nFable 5 is 88. Yes, 88. So this model is\nonly.1% behind Fable 5. And then on the\nCyber Gym, which is Cyber Security, the\nmodel actually beats Fable 5. Fable 5\nsits at 83.1 while Deepseek version 4\nPro, the new one that we got today, is\nat 83.3. Now, if this tells you\nsomething is that this model is once\nagain really geared at cyber security.\nWe saw the Flash model also be geared\ntowards cyber security and becoming a\nmodel that was quite capable in that\narea. And we're seeing the same thing\ntoday with the new version 4 Pro which\nsits at 83.3 outperforming Fable 5 on\nthe benchmark. On deep software\nengineering, this model is not beating\nFable 5 because Fable 5 sits at 70, but\nthe model achieves 62.7 which is ahead\nof GLM 5.2 and a bit behind Kimmy K3\nwhich sits at 67.5.\nBut one thing again, this is a big jump\ncompared to the preview version of the\nmodel which was at 12.8. So yeah, they\nhave really trained this model and they\nhave improved it on the back end because\nwe can clearly see in this deep software\nengineering benchmark. And there's a\ncouple of other benchmarks, but one key\nbenchmark that this model excels at is\nthe automation bench where the model\nachieves 31.8 and Fable 5 29.1. So once\nagain, the model outperforms Fable 5.\nNow this is a fraction of the cost as I\nmentioned. This is 43 cents and 87 for\ninput and output respectively versus\nFable 5 which is at $10.50.\nSo yes, if we look at the terminal bench\nfor example, we are seeing this model\nachieve a result which is 0.1 behind the\nbest model out there Fable 5 and the\npricing is insane to look at $43 versus\n$10 87 versus $50 per million input\ntokens. So when we turn that into a\nfraction, this model is 57 times\ncheaper. And earlier last week, I think\nI made a video about how DeepSeek\nversion 4 Pro was going to achieve a\nresult like this because there was a\ninvestor report that leaked. And when I\nsaw the pricing where it said that it\nwas going to match Fable 5 or even\noutperform it and be at 57 times cheaper\nper cost. When I was reading the\ninvestor report, I didn't really believe\nit. But today, it is proven because at\nleast on the benchmark so far, yes, I'm\njust saying the benchmarks, we are\nseeing similar level performance. Now,\nif this model actually performs like\nthat in production, it's a little too\nearly to tell yet. We still would have\nto give it a couple of weeks and see how\nit's performing in the long run because\nsometimes on the release day the models\nperform quite good, but over time they\nkind of deteriorate in their quality.\nSo, we hope DeepS doesn't do that.\nHistorically, they haven't done that,\nbut let's just see. Just going to put it\nout there because right now we're basing\nthis off of benchmarks. But that's it\nfor today's video. Make sure you guys\nare subscribed to the channel. Follow\nour new newsletter as well at\nuniverseofai.behive.com\nuniverseofai.behive.com\nas well as subscribe to the main channel\nWorld of AI and support us on X by\nfollowing the Universe of AIZ as well.\nUntil then, I'll see you guys in the\nnext", "language": "en", "duration": 626.126063, "segments": [ { "start": 0.08, "end": 2.639, "text": "So, Deepseek version 4 Pro is officially", "source": "source_caption" }, { "start": 2.639, "end": 4.799, "text": "out today. Now, you might be confused", "source": "source_caption" }, { "start": 4.799, "end": 6.56, "text": "because this model was technically out,", "source": "source_caption" }, { "start": 6.56, "end": 8.8, "text": "but it was in general availability, but", "source": "source_caption" }, { "start": 8.8, "end": 10.4, "text": "this is the official release, meaning", "source": "source_caption" }, { "start": 10.4, "end": 12.24, "text": "that they have trained the model a", "source": "source_caption" }, { "start": 12.24, "end": 14.24, "text": "little bit more and produced a stronger", "source": "source_caption" }, { "start": 14.24, "end": 16.0, "text": "version of the model that is available", "source": "source_caption" }, { "start": 16.0, "end": 17.52, "text": "today. Now, if I were to summarize", "source": "source_caption" }, { "start": 17.52, "end": 19.52, "text": "today's release in one simple sentence,", "source": "source_caption" }, { "start": 19.52, "end": 21.52, "text": "it is that the price to performance is", "source": "source_caption" }, { "start": 21.52, "end": 23.76, "text": "becoming a very, very important thing", "source": "source_caption" }, { "start": 23.76, "end": 25.439, "text": "because not only do we have a new model", "source": "source_caption" }, { "start": 25.439, "end": 27.68, "text": "from the Deepseek team, the SpaceX team", "source": "source_caption" }, { "start": 27.68, "end": 30.24, "text": "also dropped Grock 4.6 six. And both of", "source": "source_caption" }, { "start": 30.24, "end": 32.0, "text": "these models are competing with the best", "source": "source_caption" }, { "start": 32.0, "end": 34.079, "text": "Frontier Labs at a fraction of their", "source": "source_caption" }, { "start": 34.079, "end": 37.04, "text": "cost. Let's start with Gro 4.6. And one", "source": "source_caption" }, { "start": 37.04, "end": 38.8, "text": "thing I'm going to say is that I'm", "source": "source_caption" }, { "start": 38.8, "end": 40.879, "text": "genuinely surprised by the SpaceX team.", "source": "source_caption" }, { "start": 40.879, "end": 42.48, "text": "I'm not trying to glaze them or Elon", "source": "source_caption" }, { "start": 42.48, "end": 44.079, "text": "Musk or anything. I was just very", "source": "source_caption" }, { "start": 44.079, "end": 46.48, "text": "critical of this lab. In 2025, they", "source": "source_caption" }, { "start": 46.48, "end": 48.32, "text": "dropped Grog 4 and other models like", "source": "source_caption" }, { "start": 48.32, "end": 50.399, "text": "that, but I wasn't really, you know,", "source": "source_caption" }, { "start": 50.399, "end": 52.239, "text": "mind blown with their performance", "source": "source_caption" }, { "start": 52.239, "end": 53.84, "text": "because they're pretty they're pretty", "source": "source_caption" }, { "start": 53.84, "end": 56.0, "text": "subpar compared to any of the other labs", "source": "source_caption" }, { "start": 56.0, "end": 58.48, "text": "out there. But in 2026, it looks like", "source": "source_caption" }, { "start": 58.48, "end": 60.079, "text": "things have kind of changed a little", "source": "source_caption" }, { "start": 60.079, "end": 62.399, "text": "because Grock 4.5 was quite competitive", "source": "source_caption" }, { "start": 62.399, "end": 65.6, "text": "based off what it cost and Grock 4.6 is", "source": "source_caption" }, { "start": 65.6, "end": 67.92, "text": "actually not too bad. If we take a look", "source": "source_caption" }, { "start": 67.92, "end": 70.0, "text": "at the benchmarks, for example, if we", "source": "source_caption" }, { "start": 70.0, "end": 71.92, "text": "start with the artificial analysis", "source": "source_caption" }, { "start": 71.92, "end": 74.24, "text": "intelligence index, this model achieves", "source": "source_caption" }, { "start": 74.24, "end": 78.159, "text": "a 61 and Fable 5 is at 62. Now, what's", "source": "source_caption" }, { "start": 78.159, "end": 79.68, "text": "really important to remember is that", "source": "source_caption" }, { "start": 79.68, "end": 82.159, "text": "once again, this model is quite cheap", "source": "source_caption" }, { "start": 82.159, "end": 84.4, "text": "compared to Fable 5. This is about $2", "source": "source_caption" }, { "start": 84.4, "end": 86.96, "text": "per million input tokens and $6 per", "source": "source_caption" }, { "start": 86.96, "end": 89.439, "text": "million output tokens. In Fable 5 sits", "source": "source_caption" }, { "start": 89.439, "end": 92.24, "text": "at $10 per million input tokens and $50", "source": "source_caption" }, { "start": 92.24, "end": 94.4, "text": "per million output tokens. And this is", "source": "source_caption" }, { "start": 94.4, "end": 96.4, "text": "why you start to appreciate this release", "source": "source_caption" }, { "start": 96.4, "end": 98.32, "text": "a little bit more. It might not beat the", "source": "source_caption" }, { "start": 98.32, "end": 100.0, "text": "performance of the best models is", "source": "source_caption" }, { "start": 100.0, "end": 102.88, "text": "matching them. Even GBT 5.6 so which is", "source": "source_caption" }, { "start": 102.88, "end": 104.96, "text": "a pretty capable model and this is set", "source": "source_caption" }, { "start": 104.96, "end": 108.0, "text": "at max on the artificial analysis index.", "source": "source_caption" }, { "start": 108.0, "end": 111.52, "text": "This model achieves 61 and Grock 4.6 61.", "source": "source_caption" }, { "start": 111.52, "end": 113.84, "text": "So it ties it and it's much cheaper. And", "source": "source_caption" }, { "start": 113.84, "end": 115.119, "text": "then even on all of these other", "source": "source_caption" }, { "start": 115.119, "end": 117.439, "text": "benchmarks, for example, the GDP Val", "source": "source_caption" }, { "start": 117.439, "end": 119.6, "text": "one, it actually beats Fable 5, which is", "source": "source_caption" }, { "start": 119.6, "end": 124.159, "text": "at 1741. And then GPT 5.6, it's at 1728.", "source": "source_caption" }, { "start": 124.159, "end": 126.0, "text": "I'm not sure why they didn't choose Opus", "source": "source_caption" }, { "start": 126.0, "end": 127.52, "text": "5 as well, but I guess they wanted to", "source": "source_caption" }, { "start": 127.52, "end": 129.759, "text": "choose the quote unquote strongest model", "source": "source_caption" }, { "start": 129.759, "end": 131.28, "text": "lineup from each lab. And they chose", "source": "source_caption" }, { "start": 131.28, "end": 133.76, "text": "Fable 5 for Enthropic, which is fair.", "source": "source_caption" }, { "start": 133.76, "end": 135.44, "text": "And then Deep Software Engineering one,", "source": "source_caption" }, { "start": 135.44, "end": 137.52, "text": "which is a critical benchmark. This", "source": "source_caption" }, { "start": 137.52, "end": 140.64, "text": "model doesn't beat Fable 5 or GPT 5.6", "source": "source_caption" }, { "start": 140.64, "end": 142.48, "text": "six soul, but it gets close to it. It's", "source": "source_caption" }, { "start": 142.48, "end": 146.08, "text": "65.9 and Fable 5 sits at 70%. But if", "source": "source_caption" }, { "start": 146.08, "end": 147.52, "text": "you're getting results that are pretty", "source": "source_caption" }, { "start": 147.52, "end": 149.2, "text": "close and the model is five times", "source": "source_caption" }, { "start": 149.2, "end": 150.879, "text": "cheaper, I wouldn't be too disappointed", "source": "source_caption" }, { "start": 150.879, "end": 152.56, "text": "with this result. And then same thing", "source": "source_caption" }, { "start": 152.56, "end": 154.879, "text": "with Cursor Bench 3.2, the model", "source": "source_caption" }, { "start": 154.879, "end": 156.959, "text": "achieves 69.9%,", "source": "source_caption" }, { "start": 156.959, "end": 160.08, "text": "funny number. And Fable 5 sits at 70.5.", "source": "source_caption" }, { "start": 160.08, "end": 162.4, "text": "So once again, closer. And Grock 4.6", "source": "source_caption" }, { "start": 162.4, "end": 165.2, "text": "beats GPT 5.6 Soul. Same thing with the", "source": "source_caption" }, { "start": 165.2, "end": 168.0, "text": "Frontier Code. It gets close to Fable 5,", "source": "source_caption" }, { "start": 168.0, "end": 171.04, "text": "beats GPT 5.6 Soul. So yeah, this model", "source": "source_caption" }, { "start": 171.04, "end": 173.28, "text": "is actually available in cursor. So the", "source": "source_caption" }, { "start": 173.28, "end": 175.2, "text": "partnership with cursor or I guess the", "source": "source_caption" }, { "start": 175.2, "end": 177.36, "text": "acquisition has really helped SpaceX", "source": "source_caption" }, { "start": 177.36, "end": 179.36, "text": "make some strides in the AI space this", "source": "source_caption" }, { "start": 179.36, "end": 181.84, "text": "year and Grock build is something that", "source": "source_caption" }, { "start": 181.84, "end": 183.68, "text": "you know maybe not a lot of us have been", "source": "source_caption" }, { "start": 183.68, "end": 185.599, "text": "using so far but it's probably going to", "source": "source_caption" }, { "start": 185.599, "end": 188.159, "text": "be another platform like codeex or cloud", "source": "source_caption" }, { "start": 188.159, "end": 190.319, "text": "code that we start to use but obviously", "source": "source_caption" }, { "start": 190.319, "end": 192.319, "text": "cursor is quite strong as well. So you", "source": "source_caption" }, { "start": 192.319, "end": 194.319, "text": "have options available for you to use", "source": "source_caption" }, { "start": 194.319, "end": 196.319, "text": "them in both. And one thing to note is", "source": "source_caption" }, { "start": 196.319, "end": 198.56, "text": "that they're offering two times usage", "source": "source_caption" }, { "start": 198.56, "end": 200.56, "text": "inside Grock build and cursor for the", "source": "source_caption" }, { "start": 200.56, "end": 202.239, "text": "first week. So if you just want to try", "source": "source_caption" }, { "start": 202.239, "end": 204.64, "text": "it out, see what you feel about it, then", "source": "source_caption" }, { "start": 204.64, "end": 206.4, "text": "you know it might be worth trying it out", "source": "source_caption" }, { "start": 206.4, "end": 208.0, "text": "right now cuz you get double the usage", "source": "source_caption" }, { "start": 208.0, "end": 210.0, "text": "in the first week. Now if we were to", "source": "source_caption" }, { "start": 210.0, "end": 211.519, "text": "take a look at some of the outputs that", "source": "source_caption" }, { "start": 211.519, "end": 213.28, "text": "people have been generating with Grock", "source": "source_caption" }, { "start": 213.28, "end": 215.519, "text": "4.6, what we're looking at right now is", "source": "source_caption" }, { "start": 215.519, "end": 218.08, "text": "a Falcon 9 booster return sequence", "source": "source_caption" }, { "start": 218.08, "end": 220.0, "text": "simulation. And this was done in a", "source": "source_caption" }, { "start": 220.0, "end": 223.36, "text": "single HTML file. And as I said, if you", "source": "source_caption" }, { "start": 223.36, "end": 225.36, "text": "expected to get this type of output from", "source": "source_caption" }, { "start": 225.36, "end": 228.56, "text": "Grock in 2025, you would be kind of", "source": "source_caption" }, { "start": 228.56, "end": 230.08, "text": "surprised because you wouldn't expect", "source": "source_caption" }, { "start": 230.08, "end": 231.68, "text": "something like this to be generated with", "source": "source_caption" }, { "start": 231.68, "end": 233.36, "text": "Grock. But now it looks like we have to", "source": "source_caption" }, { "start": 233.36, "end": 235.12, "text": "start taking the Grock team a little bit", "source": "source_caption" }, { "start": 235.12, "end": 237.36, "text": "more serious because this output is", "source": "source_caption" }, { "start": 237.36, "end": 239.36, "text": "quite competitive. And obviously, we're", "source": "source_caption" }, { "start": 239.36, "end": 241.12, "text": "just looking at a simulation and we're", "source": "source_caption" }, { "start": 241.12, "end": 242.879, "text": "just basing it off of a visual", "source": "source_caption" }, { "start": 242.879, "end": 244.959, "text": "representation. But if the model is able", "source": "source_caption" }, { "start": 244.959, "end": 246.08, "text": "to produce something like this", "source": "source_caption" }, { "start": 246.08, "end": 248.4, "text": "consistently, then I would expect a lot", "source": "source_caption" }, { "start": 248.4, "end": 250.159, "text": "of people to start adopting Grock", "source": "source_caption" }, { "start": 250.159, "end": 251.84, "text": "because number one, it is cheaper than", "source": "source_caption" }, { "start": 251.84, "end": 253.68, "text": "the other labs at least at the moment", "source": "source_caption" }, { "start": 253.68, "end": 255.12, "text": "because we don't know if this pricing", "source": "source_caption" }, { "start": 255.12, "end": 257.04, "text": "strategy is going to be sustainable for", "source": "source_caption" }, { "start": 257.04, "end": 259.04, "text": "the SpaceX team in the long run. But at", "source": "source_caption" }, { "start": 259.04, "end": 260.56, "text": "least for now, their models are", "source": "source_caption" }, { "start": 260.56, "end": 262.079, "text": "definitely cheaper compared to the", "source": "source_caption" }, { "start": 262.079, "end": 264.16, "text": "others. And as I mentioned, this model", "source": "source_caption" }, { "start": 264.16, "end": 266.16, "text": "excelled at the artificial analysis", "source": "source_caption" }, { "start": 266.16, "end": 268.8, "text": "index. This model jumped to probably", "source": "source_caption" }, { "start": 268.8, "end": 272.72, "text": "number four model. It's tied to GPT 5.6", "source": "source_caption" }, { "start": 272.72, "end": 274.0, "text": "pretty much similar. So you could say", "source": "source_caption" }, { "start": 274.0, "end": 275.759, "text": "number three as well, but the models", "source": "source_caption" }, { "start": 275.759, "end": 278.639, "text": "before that are Fable 5 and Opus 5,", "source": "source_caption" }, { "start": 278.639, "end": 280.72, "text": "which are only above the model by about", "source": "source_caption" }, { "start": 280.72, "end": 283.84, "text": "a 1% or a 2% difference. So yeah, even", "source": "source_caption" }, { "start": 283.84, "end": 286.0, "text": "on this intelligence index, which if", "source": "source_caption" }, { "start": 286.0, "end": 287.52, "text": "you're not familiar with has nine", "source": "source_caption" }, { "start": 287.52, "end": 289.36, "text": "evaluations. So on all of these", "source": "source_caption" }, { "start": 289.36, "end": 291.44, "text": "evaluations, it's kind of matching", "source": "source_caption" }, { "start": 291.44, "end": 293.6, "text": "almost Fable 5 performance, which is", "source": "source_caption" }, { "start": 293.6, "end": 296.24, "text": "crazy to see. Before we continue, we", "source": "source_caption" }, { "start": 296.24, "end": 297.84, "text": "just launched the Universe of AI", "source": "source_caption" }, { "start": 297.84, "end": 299.6, "text": "newsletter. If you want to stay on top", "source": "source_caption" }, { "start": 299.6, "end": 301.6, "text": "of AI news without having to hunt for", "source": "source_caption" }, { "start": 301.6, "end": 303.68, "text": "it, link is in the description. Don't", "source": "source_caption" }, { "start": 303.68, "end": 305.52, "text": "miss out. And what you see on screen", "source": "source_caption" }, { "start": 305.52, "end": 307.68, "text": "right now is a racing game that Grock", "source": "source_caption" }, { "start": 307.68, "end": 310.4, "text": "4.6 build. And based off of this post,", "source": "source_caption" }, { "start": 310.4, "end": 312.32, "text": "the model took about 1 minute and it was", "source": "source_caption" }, { "start": 312.32, "end": 314.4, "text": "a fiveword prompt, which was a create a", "source": "source_caption" }, { "start": 314.4, "end": 316.639, "text": "simple racing game in HTML. So, if", "source": "source_caption" }, { "start": 316.639, "end": 317.919, "text": "you're able to generate something like", "source": "source_caption" }, { "start": 317.919, "end": 320.96, "text": "this easily using Grock 4.6, I think a", "source": "source_caption" }, { "start": 320.96, "end": 322.88, "text": "lot of people will be happy. And this is", "source": "source_caption" }, { "start": 322.88, "end": 325.199, "text": "a more detailed analysis of what it", "source": "source_caption" }, { "start": 325.199, "end": 327.919, "text": "costs to run GPT 5.6 6 on the artificial", "source": "source_caption" }, { "start": 327.919, "end": 329.919, "text": "analysis index and what it produced", "source": "source_caption" }, { "start": 329.919, "end": 332.0, "text": "meaning the output. Both of these models", "source": "source_caption" }, { "start": 332.0, "end": 334.16, "text": "if you remember scored 61 on the", "source": "source_caption" }, { "start": 334.16, "end": 336.24, "text": "artificial analysis index. Now to run", "source": "source_caption" }, { "start": 336.24, "end": 338.88, "text": "the whole test with GPT 5.6 it cost", "source": "source_caption" }, { "start": 338.88, "end": 342.4, "text": "about 2.8K and then with Grock 4.6 it", "source": "source_caption" }, { "start": 342.4, "end": 345.36, "text": "cost about 1.1Kish. And this tells you", "source": "source_caption" }, { "start": 345.36, "end": 346.96, "text": "that you're getting similar level of", "source": "source_caption" }, { "start": 346.96, "end": 349.199, "text": "performance at half the cost. So yeah,", "source": "source_caption" }, { "start": 349.199, "end": 351.12, "text": "this is a big release for the SpaceX", "source": "source_caption" }, { "start": 351.12, "end": 352.639, "text": "team because they just proved once again", "source": "source_caption" }, { "start": 352.639, "end": 354.4, "text": "that they are a lab that you seriously", "source": "source_caption" }, { "start": 354.4, "end": 356.32, "text": "start need to considering, especially in", "source": "source_caption" }, { "start": 356.32, "end": 358.56, "text": "2026. And I'm going to talk more about", "source": "source_caption" }, { "start": 358.56, "end": 361.12, "text": "Deep Seek version 4 Pro GA. But", "source": "source_caption" }, { "start": 361.12, "end": 362.88, "text": "basically what we're seeing today is", "source": "source_caption" }, { "start": 362.88, "end": 364.639, "text": "that both of these releases kind of", "source": "source_caption" }, { "start": 364.639, "end": 367.36, "text": "emphasize the fact that performance and", "source": "source_caption" }, { "start": 367.36, "end": 369.199, "text": "all above that is the price at what", "source": "source_caption" }, { "start": 369.199, "end": 370.639, "text": "you're getting for that performance is", "source": "source_caption" }, { "start": 370.639, "end": 372.479, "text": "becoming more and more important for all", "source": "source_caption" }, { "start": 372.479, "end": 374.8, "text": "users because we see many labs now", "source": "source_caption" }, { "start": 374.8, "end": 377.039, "text": "focusing on creating the best model at", "source": "source_caption" }, { "start": 377.039, "end": 379.6, "text": "the cheapest cost. Last year in 2025", "source": "source_caption" }, { "start": 379.6, "end": 381.68, "text": "most of the labs were just focused on I", "source": "source_caption" }, { "start": 381.68, "end": 384.4, "text": "would say creating the strongest model.", "source": "source_caption" }, { "start": 384.4, "end": 386.08, "text": "Yes, cost was important, but I think", "source": "source_caption" }, { "start": 386.08, "end": 387.84, "text": "most of the times the frontier labs,", "source": "source_caption" }, { "start": 387.84, "end": 390.08, "text": "meaning OpenAI, Enthropic, were kind of", "source": "source_caption" }, { "start": 390.08, "end": 391.759, "text": "more lenient on that fact because they", "source": "source_caption" }, { "start": 391.759, "end": 394.0, "text": "didn't have as strong of a competition.", "source": "source_caption" }, { "start": 394.0, "end": 396.479, "text": "Intelligent models that are maybe not", "source": "source_caption" }, { "start": 396.479, "end": 399.039, "text": "always ahead of OpenAI Enthropic, but", "source": "source_caption" }, { "start": 399.039, "end": 400.72, "text": "match their performance at a fraction of", "source": "source_caption" }, { "start": 400.72, "end": 402.8, "text": "the cost. So yes, price toerformance", "source": "source_caption" }, { "start": 402.8, "end": 405.68, "text": "ratio is becoming a critical I would say", "source": "source_caption" }, { "start": 405.68, "end": 408.639, "text": "indicator in 2026. Now, this is the", "source": "source_caption" }, { "start": 408.639, "end": 410.479, "text": "updated benchmark chart after the", "source": "source_caption" }, { "start": 410.479, "end": 412.24, "text": "release of the new model. And one thing", "source": "source_caption" }, { "start": 412.24, "end": 413.919, "text": "you'll see across the board is that it", "source": "source_caption" }, { "start": 413.919, "end": 415.919, "text": "matches the top level performance of", "source": "source_caption" }, { "start": 415.919, "end": 417.36, "text": "many of the models. The one thing", "source": "source_caption" }, { "start": 417.36, "end": 418.479, "text": "interesting over here is that they", "source": "source_caption" }, { "start": 418.479, "end": 420.639, "text": "haven't put Opus 5 here for some reason.", "source": "source_caption" }, { "start": 420.639, "end": 422.24, "text": "There is Fable 5 here that we can", "source": "source_caption" }, { "start": 422.24, "end": 424.0, "text": "compare this model against. But one", "source": "source_caption" }, { "start": 424.0, "end": 426.0, "text": "thing you'll notice is that Deep Seek,", "source": "source_caption" }, { "start": 426.0, "end": 429.199, "text": "remember this model costs 43.5 cents per", "source": "source_caption" }, { "start": 429.199, "end": 432.08, "text": "million input tokens and 87 cents per", "source": "source_caption" }, { "start": 432.08, "end": 433.919, "text": "million output tokens. While the other", "source": "source_caption" }, { "start": 433.919, "end": 435.68, "text": "models all over here are way more", "source": "source_caption" }, { "start": 435.68, "end": 437.44, "text": "expensive than that. So the first thing", "source": "source_caption" }, { "start": 437.44, "end": 440.0, "text": "if you look at terminal bench 2.1 the", "source": "source_caption" }, { "start": 440.0, "end": 442.319, "text": "model scores 87.9.", "source": "source_caption" }, { "start": 442.319, "end": 444.96, "text": "The older version of the model was 72.1", "source": "source_caption" }, { "start": 444.96, "end": 447.12, "text": "and the flash version which we got last", "source": "source_caption" }, { "start": 447.12, "end": 450.319, "text": "week was 82.7. And what's crazy is that", "source": "source_caption" }, { "start": 450.319, "end": 454.0, "text": "Fable 5 is 88. Yes, 88. So this model is", "source": "source_caption" }, { "start": 454.0, "end": 457.36, "text": "only.1% behind Fable 5. And then on the", "source": "source_caption" }, { "start": 457.36, "end": 459.52, "text": "Cyber Gym, which is Cyber Security, the", "source": "source_caption" }, { "start": 459.52, "end": 462.24, "text": "model actually beats Fable 5. Fable 5", "source": "source_caption" }, { "start": 462.24, "end": 465.28, "text": "sits at 83.1 while Deepseek version 4", "source": "source_caption" }, { "start": 465.28, "end": 467.039, "text": "Pro, the new one that we got today, is", "source": "source_caption" }, { "start": 467.039, "end": 469.68, "text": "at 83.3. Now, if this tells you", "source": "source_caption" }, { "start": 469.68, "end": 471.68, "text": "something is that this model is once", "source": "source_caption" }, { "start": 471.68, "end": 473.759, "text": "again really geared at cyber security.", "source": "source_caption" }, { "start": 473.759, "end": 475.599, "text": "We saw the Flash model also be geared", "source": "source_caption" }, { "start": 475.599, "end": 477.28, "text": "towards cyber security and becoming a", "source": "source_caption" }, { "start": 477.28, "end": 478.8, "text": "model that was quite capable in that", "source": "source_caption" }, { "start": 478.8, "end": 480.56, "text": "area. And we're seeing the same thing", "source": "source_caption" }, { "start": 480.56, "end": 482.72, "text": "today with the new version 4 Pro which", "source": "source_caption" }, { "start": 482.72, "end": 485.84, "text": "sits at 83.3 outperforming Fable 5 on", "source": "source_caption" }, { "start": 485.84, "end": 487.68, "text": "the benchmark. On deep software", "source": "source_caption" }, { "start": 487.68, "end": 489.919, "text": "engineering, this model is not beating", "source": "source_caption" }, { "start": 489.919, "end": 492.639, "text": "Fable 5 because Fable 5 sits at 70, but", "source": "source_caption" }, { "start": 492.639, "end": 495.44, "text": "the model achieves 62.7 which is ahead", "source": "source_caption" }, { "start": 495.44, "end": 498.639, "text": "of GLM 5.2 and a bit behind Kimmy K3", "source": "source_caption" }, { "start": 498.639, "end": 500.879, "text": "which sits at 67.5.", "source": "source_caption" }, { "start": 500.879, "end": 502.879, "text": "But one thing again, this is a big jump", "source": "source_caption" }, { "start": 502.879, "end": 504.56, "text": "compared to the preview version of the", "source": "source_caption" }, { "start": 504.56, "end": 507.36, "text": "model which was at 12.8. So yeah, they", "source": "source_caption" }, { "start": 507.36, "end": 508.96, "text": "have really trained this model and they", "source": "source_caption" }, { "start": 508.96, "end": 510.56, "text": "have improved it on the back end because", "source": "source_caption" }, { "start": 510.56, "end": 512.88, "text": "we can clearly see in this deep software", "source": "source_caption" }, { "start": 512.88, "end": 514.88, "text": "engineering benchmark. And there's a", "source": "source_caption" }, { "start": 514.88, "end": 516.719, "text": "couple of other benchmarks, but one key", "source": "source_caption" }, { "start": 516.719, "end": 519.12, "text": "benchmark that this model excels at is", "source": "source_caption" }, { "start": 519.12, "end": 521.279, "text": "the automation bench where the model", "source": "source_caption" }, { "start": 521.279, "end": 525.76, "text": "achieves 31.8 and Fable 5 29.1. So once", "source": "source_caption" }, { "start": 525.76, "end": 528.0, "text": "again, the model outperforms Fable 5.", "source": "source_caption" }, { "start": 528.0, "end": 529.92, "text": "Now this is a fraction of the cost as I", "source": "source_caption" }, { "start": 529.92, "end": 532.48, "text": "mentioned. This is 43 cents and 87 for", "source": "source_caption" }, { "start": 532.48, "end": 534.64, "text": "input and output respectively versus", "source": "source_caption" }, { "start": 534.64, "end": 537.519, "text": "Fable 5 which is at $10.50.", "source": "source_caption" }, { "start": 537.519, "end": 539.279, "text": "So yes, if we look at the terminal bench", "source": "source_caption" }, { "start": 539.279, "end": 541.279, "text": "for example, we are seeing this model", "source": "source_caption" }, { "start": 541.279, "end": 543.6, "text": "achieve a result which is 0.1 behind the", "source": "source_caption" }, { "start": 543.6, "end": 546.16, "text": "best model out there Fable 5 and the", "source": "source_caption" }, { "start": 546.16, "end": 549.519, "text": "pricing is insane to look at $43 versus", "source": "source_caption" }, { "start": 549.519, "end": 553.519, "text": "$10 87 versus $50 per million input", "source": "source_caption" }, { "start": 553.519, "end": 555.36, "text": "tokens. So when we turn that into a", "source": "source_caption" }, { "start": 555.36, "end": 558.16, "text": "fraction, this model is 57 times", "source": "source_caption" }, { "start": 558.16, "end": 560.399, "text": "cheaper. And earlier last week, I think", "source": "source_caption" }, { "start": 560.399, "end": 562.16, "text": "I made a video about how DeepSeek", "source": "source_caption" }, { "start": 562.16, "end": 564.48, "text": "version 4 Pro was going to achieve a", "source": "source_caption" }, { "start": 564.48, "end": 565.839, "text": "result like this because there was a", "source": "source_caption" }, { "start": 565.839, "end": 567.839, "text": "investor report that leaked. And when I", "source": "source_caption" }, { "start": 567.839, "end": 569.44, "text": "saw the pricing where it said that it", "source": "source_caption" }, { "start": 569.44, "end": 570.959, "text": "was going to match Fable 5 or even", "source": "source_caption" }, { "start": 570.959, "end": 573.76, "text": "outperform it and be at 57 times cheaper", "source": "source_caption" }, { "start": 573.76, "end": 575.44, "text": "per cost. When I was reading the", "source": "source_caption" }, { "start": 575.44, "end": 577.12, "text": "investor report, I didn't really believe", "source": "source_caption" }, { "start": 577.12, "end": 579.04, "text": "it. But today, it is proven because at", "source": "source_caption" }, { "start": 579.04, "end": 581.04, "text": "least on the benchmark so far, yes, I'm", "source": "source_caption" }, { "start": 581.04, "end": 582.88, "text": "just saying the benchmarks, we are", "source": "source_caption" }, { "start": 582.88, "end": 585.04, "text": "seeing similar level performance. Now,", "source": "source_caption" }, { "start": 585.04, "end": 586.959, "text": "if this model actually performs like", "source": "source_caption" }, { "start": 586.959, "end": 588.959, "text": "that in production, it's a little too", "source": "source_caption" }, { "start": 588.959, "end": 590.72, "text": "early to tell yet. We still would have", "source": "source_caption" }, { "start": 590.72, "end": 592.399, "text": "to give it a couple of weeks and see how", "source": "source_caption" }, { "start": 592.399, "end": 594.399, "text": "it's performing in the long run because", "source": "source_caption" }, { "start": 594.399, "end": 596.399, "text": "sometimes on the release day the models", "source": "source_caption" }, { "start": 596.399, "end": 598.48, "text": "perform quite good, but over time they", "source": "source_caption" }, { "start": 598.48, "end": 600.24, "text": "kind of deteriorate in their quality.", "source": "source_caption" }, { "start": 600.24, "end": 602.0, "text": "So, we hope DeepS doesn't do that.", "source": "source_caption" }, { "start": 602.0, "end": 603.68, "text": "Historically, they haven't done that,", "source": "source_caption" }, { "start": 603.68, "end": 605.6, "text": "but let's just see. Just going to put it", "source": "source_caption" }, { "start": 605.6, "end": 607.12, "text": "out there because right now we're basing", "source": "source_caption" }, { "start": 607.12, "end": 609.76, "text": "this off of benchmarks. But that's it", "source": "source_caption" }, { "start": 609.76, "end": 611.519, "text": "for today's video. Make sure you guys", "source": "source_caption" }, { "start": 611.519, "end": 613.519, "text": "are subscribed to the channel. Follow", "source": "source_caption" }, { "start": 613.519, "end": 615.12, "text": "our new newsletter as well at", "source": "source_caption" }, { "start": 615.12, "end": 617.35, "text": "universeofai.behive.com", "source": "source_caption" }, { "start": 617.35, "end": 617.36, "text": "universeofai.behive.com", "source": "source_caption" }, { "start": 617.36, "end": 619.04, "text": "as well as subscribe to the main channel", "source": "source_caption" }, { "start": 619.04, "end": 621.519, "text": "World of AI and support us on X by", "source": "source_caption" }, { "start": 621.519, "end": 623.76, "text": "following the Universe of AIZ as well.", "source": "source_caption" }, { "start": 623.76, "end": 625.36, "text": "Until then, I'll see you guys in the", "source": "source_caption" }, { "start": 625.36, "end": 627.519, "text": "next", "source": "source_caption" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "en", "duration": 626.126063, "reference_path": "/Users/sagawa/AI_WORK/video_notes/url/20260813_143826__DeepSeek_V4_Pro_Is_Out_Today_And_It_s_A_Problem_For_Every_AI_Lab/reference_subtitle.vtt", "reference_kind": "automatic_original", "reference_language": "en-orig", "selected_source": "youtube_auto_retimed_with_whisper_anchors", "decision_reason": "automatic_caption_cumulative_drift_repaired_by_monotonic_anchors", "whisper_rejected": false, "source_profile": { "segment_count": 292, "multilingual": false, "first_start": 0.08, "final_end": 627.519, "duration_coverage": 1.0, "maximum_gap_seconds": 0.08, "inverted_count": 0, "language": { "requested": "en", "matched": true, "latin": 8426, "han": 0, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": true, "usable_partial": true }, "whisper_profile": { "segment_count": 115, "multilingual": false, "first_start": 0.0, "final_end": 625.706, "duration_coverage": 0.999329, "maximum_gap_seconds": 0.5, "inverted_count": 0, "language": { "requested": "en", "matched": true, "latin": 8473, "han": 0, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": true, "usable_partial": true }, "comparison": { "window_seconds": 30.0, "compared_window_count": 21, "median_score": 0.9355, "divergent_window_count": 0, "divergent_ratio": 0.0, "divergent_windows": [] }, "timestamp_alignment": { "applied": false, "estimated_offset_seconds": 0.0, "matched_segment_count": 0, "match_scores": [] }, "caption_timing_repair": { "method": "monotonic_multi_anchor_piecewise_interpolation", "applied": true, "anchor_count": 42, "anchors": [ { "source_time": 6.16, "whisper_time": 7.6, "score": 0.884 }, { "source_time": 18.84, "whisper_time": 22.93, "score": 0.7503 }, { "source_time": 32.12, "whisper_time": 38.26, "score": 0.5769 }, { "source_time": 58.919, "whisper_time": 53.96, "score": 0.6997 }, { "source_time": 71.88, "whisper_time": 70.88, "score": 0.7914 }, { "source_time": 85.2, "whisper_time": 88.39, "score": 0.7154 }, { "source_time": 98.6, "whisper_time": 102.12, "score": 0.6634 }, { "source_time": 117.84, "whisper_time": 113.98, "score": 0.8307 }, { "source_time": 128.959, "whisper_time": 128.28, "score": 0.812 }, { "source_time": 139.92, "whisper_time": 142.59, "score": 0.6631 }, { "source_time": 153.08, "whisper_time": 155.64, "score": 0.7572 }, { "source_time": 166.68, "whisper_time": 170.6, "score": 0.6845 }, { "source_time": 179.44, "whisper_time": 184.22, "score": 0.5983 }, { "source_time": 192.079, "whisper_time": 197.4, "score": 0.6432 }, { "source_time": 218.44, "whisper_time": 212.75, "score": 0.6284 }, { "source_time": 231.36, "whisper_time": 228.07, "score": 0.8275 }, { "source_time": 243.76, "whisper_time": 243.65, "score": 0.9251 }, { "source_time": 257.159, "whisper_time": 258.31, "score": 0.9215 }, { "source_time": 271.399, "whisper_time": 274.41, "score": 0.8034 }, { "source_time": 285.039, "whisper_time": 288.9, "score": 0.6959 }, { "source_time": 297.56, "whisper_time": 302.6, "score": 0.6689 }, { "source_time": 323.279, "whisper_time": 324.2, "score": 0.8869 }, { "source_time": 336.159, "whisper_time": 338.81, "score": 0.7413 }, { "source_time": 348.4, "whisper_time": 352.7, "score": 0.7468 }, { "source_time": 373.48, "whisper_time": 367.89, "score": 0.6816 }, { "source_time": 389.2, "whisper_time": 383.75, "score": 0.6494 }, { "source_time": 401.32, "whisper_time": 400.71, "score": 0.8621 }, { "source_time": 414.639, "whisper_time": 416.38, "score": 0.8647 }, { "source_time": 427.279, "whisper_time": 431.55, "score": 0.6719 }, { "source_time": 449.84, "whisper_time": 448.643, "score": 0.8843 }, { "source_time": 463.52, "whisper_time": 463.346, "score": 0.9822 }, { "source_time": 479.799, "whisper_time": 477.266, "score": 0.8392 }, { "source_time": 492.24, "whisper_time": 491.386, "score": 0.9554 }, { "source_time": 505.76, "whisper_time": 506.246, "score": 0.9218 }, { "source_time": 519.32, "whisper_time": 521.526, "score": 0.8263 }, { "source_time": 543.519, "whisper_time": 536.726, "score": 0.6771 }, { "source_time": 555.84, "whisper_time": 552.276, "score": 0.6806 }, { "source_time": 568.8, "whisper_time": 566.826, "score": 0.8673 }, { "source_time": 582.199, "whisper_time": 581.726, "score": 0.9453 }, { "source_time": 594.6, "whisper_time": 595.746, "score": 0.8744 }, { "source_time": 606.88, "whisper_time": 610.076, "score": 0.8034 }, { "source_time": 618.639, "whisper_time": 621.716, "score": 0.6339 } ], "source_block_count": 52, "whisper_block_count": 43, "matched_block_count": 42, "case_normalization": { "applied": false, "changed_segment_count": 0 }, "source_text_preserved": true, "reason": "source_text_retimed_from_monotonic_whisper_anchors", "minimum_anchor_delta_seconds": -6.793, "maximum_anchor_delta_seconds": 6.14, "drift_span_seconds": 12.933 }, "filled_source_profile": { "segment_count": 53, "multilingual": false, "first_start": 0.249, "final_end": 626.126, "duration_coverage": 1.0, "maximum_gap_seconds": 0.249, "inverted_count": 1, "language": { "requested": "en", "matched": true, "latin": 8426, "han": 0, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": false, "usable_partial": false }, "whisper_duration_guard": { "media_duration": 626.126, "rejected_segment_count": 0, "rejected_segments": [] } } }