AI News: The Most Insane Week So Far This Year! transcript

This week was absolutely insane. I know most weeks I talk about, "Hey, there's a lot of news this week." This week there was a lot of really, really big news. Huge model releases, new product updates, lots of stuff to talk about. I don't want to waste your time. Let's dive right in. We actually got four brand new state-of-the-art models from four big foundation labs. We got Claude Fable 5.1. We got Gemini 3.8 Flash. We got Muse Spark 1.3 out of Meta.

And of course, we've got GPT-6. All huge releases. So, I actually already made videos about most of these new models. This video here, the most overhyped and underhyped new models, I really dig into Fable 5.1 and Gemini 3.8 Flash. And then this video, GPT-6 Astra is finally here, obviously digs into GPT-6 Astra. I made two additional videos this week to talk about those three different models because if I tried to deep dive on every single new model that came out in this video, it would be like an hour and a half-long video.

So, I broke them out into separate videos this week. However, the value proposition of this channel is that you really only need to watch one video a week to get a breakdown of all of the AI news. So, I am going to give you the quick TLDR of these four new models. So, let's do it in the order that they came out. First, we got Claude Fable 5.1 and Mythos 5.1, which Mythos is only available to like cybersecurity testers and stuff.

When this model came out, it pretty much crushed everything else on the benchmarks, scientific research, 52.6 over Souls 22.4, Opus's 29, Fable's 24.7, Terminal Bench 55.8% beating out everything else. On the day it came out, it was by far the best model ever. Some interesting notes about it. When it comes to price, they claim Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads. However, if we look at Artificial Analysis here, they actually do a benchmark test for cost per task, and Fable 5.1 actually came in as the most expensive model per task at $3.69 per task, where Fable, the one that supposedly 5.1 is 25% cheaper than, comes in at 314 per task.

Supposedly, this model will also refuse requests less often. It says, "We've improved our safeguards to reduce false positives, now blocking 60% fewer false positives than before." On BU & C Bench, Fable 5.1 came in second place, but it cost $4.35 to generate this SVG image. So, it is a really good model, but it is really expensive. It also took 18 minutes to generate this one SVG here. I also prompted all of these new models to do a clone of Megabonk, as I've done in past videos, just to see what sort of design aesthetic it gives these video games, but I'm going to show those after I've talked about all four models, so we can compare them side by side.

I also have like a little bit of a point I want to prove about something, and it will make more sense if I put them all side by side. Claude Fable 5.1, when it came out, the new best of the best, new state of the art. In fact, if we look at Artificial Analysis, even after the rest of these new models came out throughout the week, Fable 5.1 still leads the Artificial Analysis benchmark. This has, for the longest time, been one of my favorite benchmarks to look at, because it's sort of like a combined benchmark, where it takes a whole bunch of other benchmarks, puts different weights on them, and then tries to find like the objectively smartest model.

But, I do have some new issues with this benchmark that I didn't have before, which I'll get to later as well. But according to them, it is the new objectively smartest model, but also the most expensive model to use. Now, the next model to come out this week was Gemini 3.8 Flash. This is a new Google model, and it's designed to be pretty inexpensive and really, really fast, but still up there with the state-of-the-art models.

So, if we take a look at the benchmarks here, Gemini 3.8 Flash is pretty inexpensive, 75 cents per million token input, 375 per million output. When you compare that to Opus, they're $5 per million input and 25 per million output, and Fable is actually double that. So, Fable's $10 per million input and $50 per million output, I believe. So, way, way, way cheap compared to the most state-of-the-art models. However, this model is really good at coding.

This Deep Seek benchmark is a benchmark that people look at because it's got a pretty good correlation between the benchmark score and how good these models actually feel when you're using them to code. Gemini 3.8 Flash, 73.7%, where the previous state-of-the-art best model on Deep Seek was Claude Opus 5, which got a 74% a 0.3% difference at a giant cost savings. Now, this model is really good in a whole bunch of these other benchmarks here, but where I've noticed it being the best is actually in coding specifically.

It is fast, it is pretty inexpensive comparatively, and it's like just as good as some of the best coding models we've seen. Taking a look at the Deep Seek page here, it actually puts Gemini 3.8 Flash on top, even above Opus 5. Now, they're both showing 74%, but the average cost per task here is $2.36 for Gemini 3.8 Flash and $11.84 for Opus 5. And then Claude Fable is at 70%. Claude Fable, when that came out, everybody was blown away by how good of a model that was.

That was like the best model anyone has ever seen. Massive, massive leap. And now we're getting a model that outperforms that where the average cost is $21.63 for Fable 5 at $2.36 for Gemini 3.8 Flash. So as far as like cost to value ratio, this Gemini 3.8 Flash is like really good value here, especially for coding. Jumping back to artificial analysis here, you can see Gemini 3.8 Flash is here, 11th place. But again, this takes in a whole bunch of different benchmarks into account.

But when we scroll down to the cost per task, Gemini 3.8 Flash is way down here at 58 cents per task. If you remember, Fable 5 is up here at 314 and 369 for 5.1. So again, cost to value ratio, huge value in Gemini right now. Taking a look at Beauty Bench, this one came in a tie for third. It's a little bit wonky, like it tries to make this really hectic crazy image with crazy hair and stuff. But Fable 5.1 cost me almost $5.

This one cost me a little over 9 cents. Fable 5.1 took 18 minutes. This one took a minute and 32 seconds. Another recent update is from Artlist, and they just launched a new feature called AI Flows. This is supposed to save you a ton of time if you're frequently using AI to make content. So basically, Flows gives you this node-based visual canvas where you can connect different image, video, and voice-over models together into one reusable workflow.

So instead of rebuilding the same process every time you start a project, you just build it once, and then next time you need it, you can swap out your inputs and let it run. You can even generate multiple variations in parallel or start with one of Artlist pre-built templates and drop your own media into it. And Artlist also just added Seed Dance 2.5 to their platform, which can now generate up to 30 seconds of 1080p video in a single generation.

It supports up to 50 reference images with much stronger consistency for keeping the same character, product, or location across an entire scene. AI Flows is live now and included with Artlist Unlimited. So, try it out at the link in the description box, and thanks so much to Artlist for supporting my channel and sponsoring this portion of today's video. So, next up this week after Gemini 3.8 Flash, we got Muse Spark out of Meta.

And this one the benchmarks confuse me. We can look at Deep Suite here, which is one of the evaluations that I really, really like to see how good at coding this is. And this is showing a 75.4, which would make it the very, very top. Now, when I look at the Deep Suite leaderboard itself, it doesn't actually list Muse Spark 1.3. We've got 1.2 way down here, but number one is 3.8 Flash at 74%. So, this one being 75.4, that means it is the new best coding model ever.

And this is where when I show you all the comparisons between my Mega bunk test, it's not going to make a lot of sense. But these benchmarks are really impressive if you care about the benchmarks. It's actually Muse Spark 1.3, which is really making me question the validity of most of the benchmarks lately. Looking at Muse on Artificial Analysis, it apparently comes in at third place behind Opus 5 and Fable 5.1. It's coming in ahead of GPT-6, ahead of Gemini 3.8 Flash by a big margin.

So, this is the other benchmark that I tend to look at a lot and put a lot of value on. And this model coming in ahead of Fable 5, ahead of GPT-6, and just barely below Opus 5 and Fable 5.1 makes me really, really question this benchmark. As far as cost per task, it's pretty much in line with Gemini 3.8 Flash, 55 cents cost per task here. But when we look at B U S I bench, and again, take these with a grain of salt because this is AI as a judge, so your own visual should be like the real decider here.

This one comes in at Let me keep scrolling here. Keep scrolling. Is it even on the first page? Yeah, here we go. Muse Spark 1.3 ranks number 20th overall. It took a minute and 4 seconds to generate. There was no cost cuz right now you can actually use this model for free through OpenRouter. But this is what it generated. And if you're not familiar with SVGs, SVGs are basically images that were generated with code. So code is telling it where to draw the lines, what colors to use, where to do the shading.

It's basically a coding benchmark to see how well these models are at using code to generate an image. So this being the top model on Deep Seek and in third place on Artificial Analysis, and this is the image it generates. And then you can pair that to Gemini 3.8 flash here, which is lower on both of those benchmarks. Fable 5.1 here, which is lower on Deep Seek. It's higher on Artificial Analysis, but lower on Deep Seek, the coding benchmark.

Does this model really look like it's a lot better at coding? I mean, don't get me wrong. It's a good model and it's basically free right now, so I can't complain about that. I'm just very confused about the scoring and it's making me question all of these benchmarks like a lot. And then finally, we got GPT-6 this week or GPT-6 Astra, which hasn't fully rolled out to everybody. It was sort of just given access to like early testers.

Maybe by the time you're watching this, more people have access. I'm not sure, but we can see in their announcement, GPT-6 Astra is rolling out today to a limited set of organizations. And over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users. So they're rolling it out, but as of this recording, most people don't have access yet. I was lucky enough to get slight early access and play with it and, you know, benchmark it and do some of my own testing.

So, I will show those in a minute, but let's take a look at their page here. When I scroll down to the benchmarks, here we are at Deep Seek, 74.1%. So, again, another model coming in lower than what Muse Spark is getting. Muse Spark was over 75%. Deep Seek would have been the number one, but Muse Spark is above this one. When it comes to the Arc AGI benchmark, this one's pretty insane. 5.6 Soul, 7.8%. Arc AGI 3, 99.9%.

They've saturated the benchmark. This benchmark is useless now at this point. So, some crazy, crazy jumps in its ability to like solve problems and stuff. Again, I made a whole breakdown on GPT-6, which is why I'm kind of brushing over it quickly in this video. But, if you remember, Deep Seek, 74.1%. If we look at the Deep Seek webpage, that would put it at number one, right? That would put it above 3.8 Flash Opus 5 on this specific coding benchmark.

However, don't forget, Muse Spark 1.3 scored a 75.4. So, even better than GPT-6. However, on my Busy Bench, images that were generated by code, it's now the first place model. This is what it generated. It took 9 minutes. The price says NA because I had early access and it wasn't running through the API. It but if we look here, the estimated API equivalent cost would have been about a dollar 94 using the standard API pricing.

And again, this one scored lower on the coding benchmark than Spark. Okay, now I want to show you a comparison between all of the different Megabunk clones that I had all these models create for us. So, first up, this is the version that Fable made. Looks pretty good. You have different characters you can use. The characters actually look pretty good. This looks like a fox and the bad guys look good. You got little like bugs and monsters.

I can change the camera angle. It's a good looking game. Pretty impressed. The colors are great. There's some weird wonkiness like with these things flashing when I hover over them, but it looks pretty good overall. I like how when I switch to shotgun, the characters basically like blow up when I hit them. Nothing really to complain about on this one other than the fact that it worked for about 2 hours to build this game, used up all of my usage for the day, and then went into using extra credits, and ended up costing me like $120 to generate this game.

So, yes, it looks good, but $100 plus and it used all of my usage credits. I'm on the 20x plan for Anthropic, which is $200 a month. It used all those credits and then started using additional credits cuz I told it it can just spend money if it goes over the usage. Looks great, insanely expensive to generate. Next up, this is the version that Gemini 3.8 Flash generated. Mega Bonk 3D. Let's go ahead and use the fox here, see how this character looks.

Doesn't really look like a fox. Looks like a little wizard. But, this version looks much, much better than the version I got out of Fable 5 when I first built something like this with Fable 5. I don't know the actual cost it used to generate cuz I actually generated this one using Cursor, and it just used my Cursor credits. I didn't go over the credits. It used a very, very small percentage of my Cursor credits. Really no complaints here.

All right, now let's check out what Muse 1.3 Spark generated. This is the Mega Bonk game that Muse generated. Yeah, I got a cube with a little cylinder and I'm shooting other cubes. This is the model that is scoring number one on Deep Swell right now and number three on Artificial Analysis. This is the Mega Bonk clone that it gave me when I put it on the highest setting for Muse 1.3. So, you can see why I'm very confused by the Deep Sweet rankings.

And then finally, here's the version that GPT-6 gave me. We've got a knight, a ranger, and a mage. Let's go ahead and use the mage just because that's probably the closest comparison to the other ones. And I mean, this one to me is probably the most aesthetically pleasing. Like I just like the way the character looks. I like the way the little bad guys look that are coming at me. It looks like the most complete real game here.

And this one it generated in like 12 minutes, I think. I prompted it and it generated super quick. And it looks great. I mean, look at the little skeleton characters. Look at the little blobs that are chasing me around. There's like bats that actually look like bats. You can tell they're supposed to be bats. And as the game goes on, more and more different types of characters come in. And it just looks good. I mean, it's very, very pleasing on the eyes.

So, that's why I'm so confused when I look at Deep Sweet, like the benchmark design for coding, having it put the new Muse Spark 1.3 as the best-performing model. And same with Artificial Analysis, it's putting this model above GPT-6, above Gemini 3.8. And I think both of those models gave me both better outputs on Busy Bench and when playing around with like my Mega Bonk test. So, yeah. The benchmarks I've relied [laughter] upon the most, I feel like I can't really trust even those anymore.

I don't feel like that's accurate to my experience anymore. I'm curious what you think about that. But again, I made two other videos this week that go even deeper on all of these models. One that specifically focuses on Fable and the new Gemini model, and another one that focuses just on GPT-6. I didn't make a separate video about Muse Spark, but I feel like I accurately conveyed my thoughts about it in this video. Anyway, I got more to talk about.

This next one is a super quick one, but one that I'm really, really happy finally exists. If you use the Chat GPT app and you want to connect more than one Google account, you can finally do that. Now, this is news that came out last week and somehow I totally missed it. If I open the Chat GPT app here, click on plugins, and then click into Gmail, I can click manage, and you can see I've got one email connected, but I can connect another account now.

I know I'm not the only one that's excited about this. Now, Google also rolled out a sort of quality of life feature for Gmail, Docs, and Google Keep. You can now use voice directly inside of those platforms. So, pretty much the Gemini audio features that you've had access to in some of the other Gemini specific apps, you can now use in Gmail, Docs, and Keep. >> Imagine you can have a conversation with your inbox. I'm a mom of a 3-year-old.

Jack is starting pre-K. What do I have to bring for Jack's first day of pre-K? >> Two changes of weather-appropriate clothes and a family photo for the classroom tree. >> I'm a person who thinks out loud. Now, I have this partner in Gemini where I can brainstorm right inside of Docs. Hey Gemini, help me create a document to help me organize my upcoming block party. >> I've gathered the details for the end of summer block party scheduled for August 29th, 2026. >> So, it says these new conversational features are rolling out this week, so you may have them in your account by the time you see this, but they are for AI Plus Pro and Ultra subscribers.

So, you got to be on like a paid plan to be using these. We got an update in the world of AI video slop, I guess. That's not my words. I mean, like the people doing this stuff are actually calling it slop. MiniMax rolled out a new video model that can generate a 15-second video in 13 seconds. So, it can make a video that is longer than the amount of time it takes to generate that video. So, of course, what do you think people did with that?

They started making infinite streams where you could just keep on prompting stuff and the video just keeps on going forever because it's generating faster than it's actually playing back. The company Foul created foul.live or fall.live, never really know how to pronounce that, but it's a new platform for infinite interactive AI live streams. You pick a channel, prompt what happens next, and watch it generate in real time.

Peter Levels here created something similar, infiniteslop.ai. 37,000 people were watching infinite slop on one of the days. But it's the same kind of idea. People tell what they want to see in the comments and the video just keeps on generating and making new videos. So, here's what that looks like on infinite slop. You can see what people are prompting and you can see what it's generating and the queue for prompts that are going to come up and it just keeps generating videos of slop. >> It's slopping time. >> Yeah, and there's 75 people watching and over 5,000 have watched today.

So, yeah, that seems like a pretty good use of your time, I guess. I mean, just because we can, does that really mean we should? Uh, anyway, I said there's a lot this week. There's still quite a bit more I want to talk about, but I will talk about all the rest very rapidly in a rapid fire. I'm going to start with one of the cooler things I saw this week, which is the new Atlas model out of World Labs. This is really cool cuz you give it an input image and you tell it a camera path and it will generate an output based on that.

So, basically using photos, it can sort of figure out the space around those photos. Atlas generates images and videos from one or more images with pixel-perfect camera control. It's got spatial reconstruction. It reconstructs real-world scenes from one to dozens of input images. Now, this isn't just generating a video. Like, this isn't a video generator. This is generating like an environment here. We can see the input image that was given, and then it sort of reconstructs everything based on that one image.

In fact, here's another example where it took this one image, and I can actually I'm manipulating this in real-time. As I move my mouse around, look at different angles from just that single image. And if you give it two images, you can see that it can combine those two images into a single scene that you can move around and manipulate. This one here looks like it's got 1 2 3 4 5 6 7 different inputs to generate this output here.

And again, [clears throat] this isn't just a video. This is like something that you can move around in and manipulate and change camera angles on and things like that. Super, super cool. Now, this is in early access, so most people don't have access to it yet. I reached out and asked about early access so I can play around with it and test it and make another video about it because I thought this was really cool. I love stuff like this.

I can imagine, you know, going and taking a whole bunch of pictures inside of your house, putting them into a platform like this, and easily remapping the house. Now, I know you can already do that with things like Gaussian splats, but this does it with a lot less photos, it seems. Something I'm really excited to play with. So, again, not a video generator, like an actual environment generator. You can also do really cool like bullet time type effects with it.

So, if you did want to use it for video, there's some pretty cool implications there as well. Anyway, this is supposed to be a rapid fire and I'm nerding out about this too much. So, moving on. The company Runway introduced something called Solaris this week, which appears to let you like manipulate videos in real-time. This is another one that's in early access, and I don't have access yet. But, check out some some demo examples here.

This person takes a selfie here, uses that photo, puts them in a scene, and they're able to drag and drop shoes on their feet. They're able to drag and drop different clothes. Here's a scene inside a room where they're manipulating it and moving things around in the scene. And what's crazy is like as they move the plants around, you'll notice the shadows change and things like that as well. So, it's built on the Gen 4.5 video generation model, which was adapted to understand interaction and respond in real time.

Solair treats user input as conditioning for the next frame, the same way it treats text or images. So, the model observes clicks, drags, and other interactions as it generates, using them as signals for what comes next. Again, this is something I probably would have done a deeper dive on if we had access to it, but if you want to check out some of the demo videos and see what this is capable of, all the links from this video are mentioned down in the description.

If you're a user of Open Claw, they just released a brand new overhauled a new version with Open Claw 2.0. For me, things like Codec and Cursor have completely replaced Open Claw since they have sort of like agentic features built in now. So, I haven't used Open Claw in quite a while, but if you are somebody that's still using Open Claw and you've tried the Open Claw 2.0 and really, really like it, let me know in the comments cuz maybe I'll make another deep dive video about Open Claw.

We got not one, but two new transcription models this week. One out of Meta, Muse Voice Transcribe. It does exactly what you think it would do. You speak, it transcribes. These transcription models are getting better and better. This one is really good for streaming, I guess for like real-time transcriptions. This one came out on September 1st, and then just 2 days later on September 3rd, Microsoft released MAI Transcribe 2, which is now the fastest, most accurate, and cheapest speech recognition model.

So, again, another model that does exactly what you'd expect it to. It transcribes your audio into text for you, and they're just getting better and better. This would have been a much bigger story if I hadn't already went off on a big long rant in last week's news video. But, Nvidia officially acquired Hugging Face. We talked about it last week because there was rumors that it was going to happen, so I really deep dove on the implications of that.

But, to sum it up, this is basically Nvidia betting on open source. You have companies like Meta and OpenAI and Google all trying to build their own chips so that they don't have to use companies like Nvidia and buy all their compute from them. So, Nvidia goes, "Okay, well, let's really pump some investment into open weight models, the models people are just going to run locally, the models that are very inexpensive to run.

And if the big frontier labs don't need our compute as much anymore, well, then the general consumers and the businesses and the enterprise companies that want to build their own on-premises sort of data centers, they'll buy our compute and use the open weight models." In my mind, that is sort of the play that Nvidia is going for here by acquiring Hugging Face. Hugging Face is like the GitHub for model weights. Whenever new models come out, people load them onto Hugging Face.

Hugging Face also is a compute provider. You can go and use the models using Hugging Face's compute. So, very interesting dynamic here, but again, this is Nvidia betting that open weight models are going to get bigger and bigger and more popular in the future, and they can sell hardware to the people that need those. But, if you want an even deeper dive, again, last week's Friday AI news video dove pretty deep into this.

Apparently, your ChatGPT conversations could end up in court, and they are not protected. So, the conversations that you are having with ChatGPT, they're not as private as you think. If you're talking about illegal stuff and you ever get in trouble, your chats are not off-limits in court. So, just something to keep in mind for the future if you're talking about some weird stuff with your models. The mayor of New York, Mayor Mamdani, made a new AI policy this week making it so that students in K through eighth grade can't use AI in school.

Now, it's not against the rules for kids to use AI, they just can't use it in school. The whole thinking behind this is that kids need to learn fundamentals without using AI before they get access to AI. So, in the same way when you're learning the basics of math, like addition, subtraction, multiplication, division, like the basics, you really shouldn't be learning with a calculator. But then once you understand how to do it, then you bring the calculator to the mix so that you can do it easier and faster.

That's sort of the thinking here, but there's obviously a lot of mixed opinions on how people are seeing this right now. And finally, I always like to end with tech and gadgets, and the company Dyson, yes, the company that makes the vacuums and the fans, they just made an AI toothbrush. And to be honest, I kind of want one. >> [laughter] >> It's called the Dyson Cam Jet, and it's got a camera and a water jet on it, and it uses AI to accurately shoot water between your teeth and like does the flossing for you.

So, you're brushing and flossing at the same time, and it uses AI to like help shoot the water jets in the optimal spots to do the flossing for you. That's my understanding at least. So, here's a picture. You've got like a camera sensor thing here. You got your normal toothbrush, and it's got a light so the camera can see, and then it's shooting the water to do the flossing. And like if I can knock out both brushing and flossing all in one go, and it uses AI to make sure it's doing a good job, like I'm all for that.

The one major downside is that it is a $500 toothbrush. >> [laughter] >> So, no, I haven't ordered one yet, but I'm kind of seriously considering it. Would you pay $500 for one of these? Anyway, that's what I got for you today. I know there was a lot this week. All of these model drops are getting sort of exhausting. If you feel exhausted by it, trust me, that's how I feel as well trying to keep up with it every day.

But that's what I do here. That's what I'm trying to do every single week. I am keeping up with all the AI news. I'm drinking from the firehose. I'm testing the things. I'm reading the articles. I'm talking to the people. I'm, you know, trying to get early access wherever I can and trying to stay as looped in as possible. I'll sort of take on that overwhelm and then make these Friday videos for you. So that if you want, you can just watch one video per week and stay completely looped in on everything that's going on in the world of AI.

That's my goal with these types of videos. If that's something that interests you, maybe consider liking and subscribing to this channel. That'll make sure videos like this continue to show up in your YouTube feed. I don't think any of the news is slowing down anytime soon. We still have Meta Connect coming up this month. We have an Apple event coming up this month. Open AI's Dev Day is still coming up this month. The Made on YouTube event is coming up this month.

There's a lot still happening. People call this Tech-tember because there's so many different like keynotes and things going on this month. I devote pretty much my full-time job's worth of attention to this stuff so that I can try to make it easy and accessible to as many people as possible. And I absolutely love doing it even though it is overwhelming at times. It's a lot of fun, but it's also a lot to keep up with.

But I wouldn't have it any other way. Anyway, thanks so much for hanging out and nerding out with me today. Hopefully you learned something. Hopefully you feel more looped in. And uh yeah, I really hope to see you in the next video. I got some really cool ones coming up. So see you there. Thanks again. Bye-bye. >> [clears throat] >> I am the knight of the round table. >> And that's how it happened.