Transcription
HERE'S the AI news you probably missed this week.
Starting with the thing that most people have been talking about lately, and that's that we finally got our Fable 5 back. Now, if you haven't been paying attention, Fable 5 was released on June 9th, 2026, and it was pretty much the best model anyone has ever seen, including me. It worked really, really well.
Then, just a few days later on June 12th, the US government basically said, "No, this model is too dangerous for it to be out in the public." Apparently, some people at Amazon had found some vulnerabilities with it, reported them to the government, and got Fable shut down. Fast forward to this week, 2 and 1/2 weeks later, and on July 1st, they finally gave us our Fable 5 back.
Now, the version that they redeployed isn't exactly the same version. Now, supposedly has all of the same capabilities, but they did have to build in some extra safeguards. According to their article, the new classifier comes at the cost of flagging benign requests more often during routine coding and debugging tasks. Now, if you remember from the last time they launched this product, the biggest complaint was that it would not actually do too many of the benign things that people asked it to. Like really simple stuff. It would say, "Sorry, due to cyber security reasons, we can't answer that." And it would push you to using one of the Opus models instead. Well, this new version apparently is going to not answer even more of your benign requests. They've got this handy little chart here to show you this that benign stuff is going to fall within this larger safety margin and actually not get sent to Fable and we'll probably just get redirected to Opus.
According to this Bridgemind account over on X, Fable 5 came back nerfed. They ran it on their own benchmark and they said that debugging went from an 86.2 down to a 25.9. Refactoring went from a 73.6 to a 38.4 and Hallucinations went from 75.9 to 61.7. They do go on to say that when the tasks are able to be completed by Fable 5, it functions the same as it was before, but the new guardrails are so strict that it gets through way less. And a lot of the comments kind of debate this. Most people saying it actually feels kind of the same. And honestly, so far from my experience, it's yeah, felt pretty much the same. I haven't actually gotten blocked on any of my own tests yet, but I'm also not trying anything that I feel like should get blocked. So, none of my benign requests are getting blocked.
I mean, check out some of this stuff that I've been able to do here. It's still really good at making games. I made this whole Cube Basher game and actually continually improved it more and more. It's very much a Mega Bonk clone, but it works really well. And the graphics are surprisingly impressive. like I can change my camera angles and everything just looks good and works good.
I also used Fable to completely build me a dashboard for creating short form videos. So I can come up here, click new short. I can give it a URL to an article. I can copy and paste text. I can drag and drop a file here. I can give it directions on how to steer the short. I can choose my target length. I can tell it whether this short is actually a promotion and give it details about the sponsor. I can tell it if this short includes an interview. And I can select any model I want for it to actually generate the script for this short width. It generates the whole script here. This one happens to be based on an interview. It fact checks a lot of what I say in this script to make sure that I'm not saying anything that's not true. I can give it feedback and it will optimize and edit the script for me. It has the formula that it used to write the script. I can export it as a doc so I can share it with my team. And it even built a teleprompter for me, which I don't really need cuz I've got a teleprompter. that it built one for me in here without me even asking it to. But then it also goes and generates B-roll for me. So if I click on B-roll here, it actually uses Remotion behind the scenes for the entire video, generates B-roll to go along with the script. So all I have to do is go and record the script that it gave me into my teleprompter here. And I have A-roll. I go and download the video clips that it generated for me here. And I overlay this video as B-roll over the A-roll that I recorded with my teleprompter. So, it pretty much does most of the work of generating shorts for me.
Now, this isn't an app that I'm going to be sharing and making publicly available. And the main reason is I designed it so that it literally watches my shorts. It watches my Instagram reels and it watches my YouTube shorts, watches for which ones are performing well and get a lot of views and then constantly optimizes to make the scripts better and better based on what was working. Now, I haven't tested making any videos with this yet, but you get the idea here. makes the script. It makes the B-roll. Tells me where to put the B-roll. And once it's done all that, I can literally download each little video here by coming up to the top and clicking download all. And I can see exactly how this whole thing got produced. This is where my cost came from. So, I have an audit log of every single video that I make. This was built with Fable and it was built within the last 24 hours or so since this new version's been live.
I also built this app that generates B-roll for me for long form videos. I basically plug in an article here. Like for instance, this article about redeploying Fable 5. Let's paste this in here. Open the article. And then I can do things like highlight sections of the article. Click save selection. Let's highlight this. Click save selection. And I'll go through and highlight a bunch of stuff. I can even come down here and pick images and highlight an image and save that to the selection. And then if I click capture page, it creates this video for me where it actually highlights the things that I selected when I was highlighting it originally, but it does it like in an animated form. So it goes through and all of the various sections, it's highlighting them. And you'll see the animation will scroll down the page and highlight various things. And then if I skip ahead in my video here, you'll notice that at the end I selected an image and it sort of like zooms in on that image. So, it allows me to talk about news articles, go and highlight different things on the news article, and then anything that I highlighted creates video B-roll for me that I can come up here and export.
I built this with Fable, but probably my crowning achievement, the thing that I'm the absolute most excited to share with you is the brand new benchmark that I created. I teased it in my previous video called Beautybench. It's a benchmark to test how well the various AI models are at drawing SVG images of Gary Buucy. And it is so much fun to go through and see the evolution of how these models were able to draw SVGs. Now SVGs are not like an image generator like Nano Banana or Dolly or Stable Diffusion. It's actually using code to draw these. So, if we go all the way back to March 2023, this is what GPT 3.5 Turbo thought that Gary Buucy looked like. This is roughly the model that the original Chat GPT was using. Chat GPT came out in November of 2022. So, this Turbo model is slightly more advanced than that, but it's roughly the original Chat GPT model. And then over time, you can kind of see how models evolved at being able to generate SVGs. And some models haven't evolved at all. But when we get up to the top here, you could actually see some pretty big jumps. So by February of 2026, Gemini 3.1 was drawing SVGs that looked like this and like this. Gros actually, for some reason, look more like doodles than they do sort of lines and circles that SVGs are known for. This entire benchmark and website I created using Fable.
Here's what GPT 5.5 and 5.5 Pro think Gary Buucy looks like in SVG form. And here's what Fable thinks that Gary Buucy looks like in SVG form. I'm partial to this Fugu model here, but I also made it so you can sort by cost, highest to lowest. So, GPT 5.5, it cost me $4 to generate this SVG with 5.5 Pro. 01 Pro cost me $3 to generate that. We can sort by tokens used, which ones used the most tokens. Grock 420 used 106,000 tokens to get me this. We could sort by run duration. Which one took the longest to generate? GLM 5.1 took 6 minutes and 50 seconds to generate this SVG here. I also built a timeline page. So you can pick any sort of model provider here and see the evolution. So if I click on OpenAI, it'll filter it down. And we can see it starts here with GPT 3.5 Turbo. And over time, you can see the improvements that the OpenAI models have done at making these SVG images all the way up until we get to GPT 5.5 and 5.5 Pro. So, this is the SVG of today. Well, you know, the latest model that came out in April of 2026 versus what the original chat GPT would have generated. I think my favorite one to look at is Meta cuz well, they just kind of never figured it out. Although, I really like this image here that Quinn generated, and it cost a little over two cents to make. And while I was at it, I figured I might as well plug in the image generators as well. This one isn't as built out as the SVGs, but we can see how the various image generators have evolved over time. Starting with MidJourney V1, which came out in February of 2022, all the way to what we're getting now from Nano Banana 2 Light and GPT Image 2, which I think is currently probably the winner. But again, there's a lot of models I haven't added in here yet. We've still got like Crea models, ideoggram models, stable diffusion models. So, this one isn't quite as fleshed out as my uh Buy Bench SVG test here.
Why did I make Buucy Bench? Not because I thought the world needed it, but because it was just so much fun to make. Like, I was just playing with Fable and adding features and testing new models. And I used Open Router to go and just bulk generate with tons and tons and tons of different models. And I lost, you know, 2 days of my life building beauty bench just because I was having a lot of fun. And sometimes just having fun with these AI models is uh a good use of time. Like what do they say? Time you enjoy wasting is not wasted time or something like that. That's why I built beauty bench. But if I'm being honest, I actually think it's a fairly legit show of how these models are improving. It's basically showing us, is this model decent at writing code that can look like faces? And you can definitely see a very clear progression over time that gives you an idea of how these models are improving. But man, has it been fun to make.
Anyway, back to Fable 5 here. One thing that is very important to note is that if you're on one of the paid plans, you get access to Fable 5, but only through July 7th. So, by the time you're watching this, you probably only have like a couple days left to get in and use Fable as much as you can on your current account. After that, it's going to be available via usage credits. You're going to have to pay extra on top of your existing plan in order to get access to Fable. So, they're not giving us a huge window to use this. So any big projects that you have in mind, it's probably a good idea to give them to Fable, have Fable work on them and do like the bulk of the work now and then when you lose access to Fable, go back to your favorite model of before, you know, one of the Opus models or GPT 5.5 and let it make the more incremental improvements, but let Fable do like the big portion of the work now. Anyway, that's what I'm doing with a lot of my projects that have been on my list.
Another cool AI product I've heard a lot about lately is GenSpark. And it's basically an all-in-one AI workspace that brings together a bunch of top AI models and productivity tools into a single platform. And one of their newest features is called GenSpark Claw. It's basically a ready-to-use version of OpenClaw that runs in the cloud. You can connect it to tools like Slack, Teams, WhatsApp, Telegram, and more, and then just tell it to do tasks for you. They also have a desktop version that can actually interact with files on your computer and navigate the web on your behalf. Another feature I thought was really interesting is their upgraded AI office suite. They've got AI slides, AI sheets, and AI docs. For presentations, there's this new guide mode that asks questions before building your deck. And if you want to change something, you can literally click on an element and tell it what to edit while keeping the rest of the presentation consistent. The AI sheets can pull data from spreadsheets, databases, or online and automatically generate formulas, charts, and visualizations for you. And AI docs can turn pretty much any word dump into a polished, professionally formatted document. Paid users currently get unlimited AI chat and AI image generation throughout 2026, subject to abuse and guard rails, of course. And they're also doing a get started bonus where new users can try their premium features for free. So, if you're interested, check out the link in the description. Users who sign up through that link get free credits to try it out. And thanks so much to GenSpark for supporting my channel and sponsoring this portion of today's video.
Now, let's move on to some other news that came out late last week that is very along similar lines to what we were just talking about with Fable, and that's the new GPT 5.6 models. Now, we actually talked in last Friday's video about how OpenAI is going to be releasing them, but most people aren't going to get them because the government wants to approve them person by person or company by company or I I don't know how that's going to work. But GPT 5.6 was coming out, but we normal people weren't going to get to use it. Well, that came out on Thursday and then OpenAI made that official on Friday, basically saying, "Hey, our next generation models are here, but you can't have them."
Now, this model comes in three flavors. You've got your Soul, your Terra, and your Luna. And it kind of sounds like maybe they're going in the direction that like Anthropic went in where like their biggest models were Opus and their medium models were Sonnet and their lowest models were Haiku. And I mean now they have Mythos class on top of that. But it kind of feels like they're going in that same direction. Like Soul is maybe their Mythos class or Opus class. I'm not quite sure. Somewhere between Mythos and Opus. And then you have Terra, which is kind of like their Sonnet. And then you have Luna, which is like their Haiku, their smaller models, but this Sole model does appear to be fairly on par with what you get out of Fable. This is like their version of Fable or Mythos or whatever you want to call it. But the big difference here is the price. Input $5, output $30. If we compare that to Fable, Fable's input is 10 and output is 50. So half the price for input and almost half the price for output. But if we look at the benchmarks that they share here, this GPT 5.6 Six. Soul Ultra, which I guess that's probably a better way to look at it. Soul Ultra is like their Fable, where Soul is like their Opus. But Soul Ultra on Terminal Bench scores a 91.9% where Fable's all the way down here at 84.3%. And their Terra, the sort of middle Sonnet version, is tied with what Fable is doing on Terminal Bench specifically.
Now, we don't really have a lot to go off of for this model yet. They haven't given us a ton of information other than some of the pricing here. Artificial Analysis doesn't have it within their benchmarking yet at all. So, nothing there. And again, their website is leaning in heavily on showing off Terminal Bench, one specific benchmark about how well it runs terminal commands essentially. But when we scroll down to the availability and pricing section during the preview, 5.6 models will initially be available through the API and codecs to a select group of trusted partners and organizations. We plan to make them more broadly available to people using chat GPT codeex and the API soon. Still no exact indication of when we're going to get access to it, but my guess is we'll probably get it within the next couple weeks. And to be fair, I wouldn't be shocked if you know that July 7th deadline where Fable is going to cut off on you. Would be kind of interesting to see OpenAI turn on their model for everybody on that date. Like, come on. That sounds like a very OpenAI thing to do, doesn't it?
Now, while we're on the topic of Open AI, I want to talk about something really, really quickly that's in the same vein, but it's also something I don't quite know how I feel about. Apparently, Open AI has been floating the idea of giving the Trump administration a 5% cut of the AI boom, or basically saying, "Hey, we'll give you 5% of Open AI." OpenAI has floated giving the US government a 5% ownership stake as a way of easing tension with the Trump administration and blunting mounting public backlash against AI. Sam argued that giving the public a financial interest in the company would be the best way to share the upside of AI. That 5% would represent roughly 42.6 billion.
Okay, so here's why I have sort of mixed feelings about this one. And I guess mixed feelings isn't even the right way to put it. This actually feels off to me. So these AI companies, they need to get regulated by the government so that the models don't become harmful, but they also need the regulations to not sort of slow down their business interests like what we've been seeing with 5.6 and Fable. But if the government has a stake in the company that the government is also in charge of trying to regulate, doesn't that create some pretty massive conflicts of interest? Because now the government stands to make more money based on the success of these AI companies which would then sort of incentivize them to let some of the regulatory hurdles slide so that models can get pushed out faster so that the government could make more money off of them. Like how's that going to work? And also the way this article was worded is the Trump administration. Shouldn't it be like they're giving it to the government so the government could give it back to the people? I don't know. It all feels very odd to me and uh yeah, I'll just get back to sharing other AI news, but this was very interesting and weird to me because I don't know how this plays out.
But let's talk about another new AI model that came out this week, and that's Claude Sonnet 5. Now, typically when a brand new Open AAI model or a brand new Anthropic model or a brand new Google model comes out, it's typically the biggest news story of the week. But this one's been totally overshadowed by everything else we just talked about. Fable getting shut down, Fable getting brought back, GPT 5.6 getting announced but not released. And it doesn't even feel like Anthropic cares that much about Sonnet. Like here's the benchmarks. Agentic coding, Sonnet still less than Opus. Multi-disciplinary reasoning still less than Opus. Computer use still less than Opus. And knowledge work slightly improved over Opus. Apparently, it's got a lower rate of undesirable behaviors than Sonnet 4.6. The main benefit of Sonnet is that it's going to be cheaper to use than Opus and Fable, but it's not going to be as good as either one. But if you need a model that's less than those and almost as good as those, then that's where Sonic comes into the mix. Even on Anthropic's system card, literally on page two, it says Sonet 5 is our most capable Sonic class model, but it is not at the capability frontier compared to more capable Opus or Mythos class models. You don't even have to get too deep into their system card for them to be going, "Yeah, it's a new model, but it's not as good as their other models."
But again, the cost is where it really matters. If you're using the API and not using like a pro subscription, you can see Fable's $10 input, $50 output for your tokens with the API. Opus is $5 in, $25 out, so half the cost of Fable. And Sonnet 5 is $2 in and $10 out. So, you know, two-fifths the cost, not even half the cost to jump to Sonnet. But that's only through August 31st. The price jumps to $3. So, three-fifths and $15 input after September 1st. But again, if you're using like a Claude subscription, you're probably just going to still use Opus. But I also know what you're wondering from here. How does it do on Beautybench? Well, you probably already saw, but let's look again. This is what Sonnet 5 did on Beautybench. In my opinion, Fugu Ultra does better. And I think even, you know, Quinn 3.7 here and Gemini 3.5 flash here do a better job. But yeah, there's your Claude Sonnet 5 beauty.
We also got some updates out of Google this week. They released a new version of Nano Banana called Nano Banana 2 Light. And they also released Gemini Omni Flash, so a faster probably not quite as high quality version of Gemini Omni for video generation and editing. So Nano Banana 2 Light is basically Nano Banana 2, just way faster and cheaper. It generates image outputs in 4 seconds and it cost about three and a half cents per 1,000 images. So really, really cheap to generate tons and tons of images. Now, it does say it's coming to the consumer platforms like AI mode and search, Gemini apps, notebook LM, Google Photos, etc. If I jump over to my Gemini account, I do appear to have access to it. I'll just do a wolf howling at the moon and it took about 6 seconds and it's pretty solid. Now, keep in mind I am on an ultra plan, so I'm not 100% sure if it's rolled out for everybody just yet. A monkey on roller skates, Gary Buucy's face, a sign in a store window with three products, their description, and their price. This one I feel like took like maybe 2 seconds longer. It took about 8 seconds instead of six, but all of them were really, really fast.
And then we have Gemini Omni Flash. Now, this is one that I've had access to and I've even made a video about it back during Google IO, but it looks like now it is rolling out to developers via the API and Google AI Studio and it costs 10 cents per second of video output. If we head over to Google AI Studio over on the right, we make sure that we are on video Gemini Omni Flash Preview. Again, this is not necessarily free to use. So, if you have like payment details on file so that you can use the API, you should be able to test this. I'm actually going to take this video that I recorded at Google IO looks like this. And we will toss this in right here and I'll say make Godzilla enter the scene and start stomping around. And well, I've tried multiple times to get it to work and it keeps giving me an internal server error. I thought maybe it wouldn't do it because it was Godzilla. So, I just said make it nighttime and it still wouldn't generate. So, I don't know, maybe it's having some issues at the exact time of this recording. However, I do have access to Omni inside of Gemini. I don't know again if it's just because I'm on the Ultra plan or if it is rolled out in Gemini now, but let's see if it works here. And yeah, I got it to work in Gemini. So, let's take a peek at our new video. We got two Godzillas. We got one in the background and one in the foreground. But you get an idea of what you can do with Omni. It's really good at taking existing videos and then sort of altering those videos. It can also generate videos, but if I'm being honest, I don't feel like it's quite as good at generating from scratch as something like Seed Dance or even the VO models, but it's okay at that as well.
And one last cool feature that Google rolled out inside of Notebook LM, you can actually create short form videos in Notebook LM now that are like 60-second vertical videos. So, if I go to my Notebook LM account here and I go to my birds aren't real notebook, which I've played with quite a few times in past videos, you can see we've got try new short video overviews. So, you can either try it here or I can just click on video overview and select short as my option. We'll go ahead and generate this. This thing better be worth it. It's been generating for like a half hour. Okay, it's finally ready. So, let's see what we got here. >> Why do people increasingly mistake obvious satire for actual facts? It's due to a mental shortcut called the bandwagon heuristic. Social media algorithms trigger this glitch to actively hack your ability to recognize a joke. Take a ridiculous spoof headline like a politician proposing mandatory vasectomies. Normally, extreme absurdity acts as a natural defense mechanism. That ridiculousness acts as a cue, signaling your brain to reject the claim as a joke. For how long that took, that was not worth it. I mean, the audio was kind of interesting, but the videos were very, very, very basic, especially since it took like at least 30, 35 minutes. I literally walked away and had to come back. It was taking so long. But anyway, you can make shorts now in Notebook LM.
All right, I know I've covered a lot and sort of rambled and ranted quite a bit in this one. But I have a handful more things that I want to share that came out this week, but I'll just run through them quickly in a rapid fire.
Starting with a new platform that came out of Anthropic called Claude Science, which is an AI workbench for scientists. So, this is like a new app sort of like Claude Code or Claude Co-work, but it's specifically designed for scientists and, you know, scientific use cases. And you can use this app if you're on one of the paid plans. Just go to claude.com/product/cloudscience and you can download it on Mac or Linux right now. I'm sure it'll be available for Windows eventually, but since I'm not really using Claude for scientific research, I, you know, haven't downloaded this or played with it myself.
If you're into coding with AI, Cursor just got an iOS app. With the iOS app, you can launch always on agents in the cloud or control agents running on your computer from your phone. I've started to do this a lot with like the Codeex app because I use Codex quite a bit. and I will have codec projects running on my computer, but I'll be controlling them and checking in on them from my phone while I'm like in the other room or while I'm away from my house. Pretty cool. So, you can do that with cursor now as well.
Also, news that some coders would appreciate. X now offers an MCP server to make it platform easier for AI tools to use. So X already had an API that developers could tap into to pull real-time data from X, but the MCP server just makes it a lot easier and a lot simpler for you to like get your agents and things like that talking to the X platform.
If you're on one of Google's paid plans like the AI Pro or Ultra subscribers, you can now have Google take notes for you when you're on Google Meet. So Gemini transcribes the conversation and creates meeting summaries with key action items. The notes are automatically saved to your Google Docs in Google Drive. And after the meeting, you'll get an email with the summary and action items. I personally tend to use Granola for this right now. The exact same thing, but now it's going to be native directly inside of Google Meet.
Gemini Spark, which is like Google's answer to OpenClaw, is now available on the Mac OS launch, but only for Ultra subscribers right now. So, I happen to have an Ultra plan and I have the Gemini app. So, if you're curious what that looks like, you can open the Gemini app, click on Spark here, click use Spark for Mac, and now you can have it control folders and do different things on your computer for you. So, if I was to add, for instance, my downloads folder. This is my example I always use, but let's go ahead and allow this. I could tell it to organize the files in my downloads folder into the already created folders. So, my downloads folder looks like this. It gets out of hand quickly, but I do have folders already created here. If I submit this, let's move my folders back over here. We should actually see it in real time start to organize my downloads folder for me. Let me just move this over here. And yeah, I mean it didn't really do it in real time or you saw it. It was just like one fell swoop. It went boop and everything just got put into the folders. But it organized my downloads folder for me. So now Spark can take actions on your behalf just like if you connected a folder in Codeex or Claude Code or Cloud Co or if you were to ask your OpenClaw or Hermes agent to do it for you. Well, now Gemini Spark can do it too in the Mac OS app.
And finally, OpenAI is teasing some new hardware for codecs. We don't have much to go off of, but we do know they have some sort of hardware coming out. It looks like some sort of macro pad where you can have like pre-programmed keys. Like I personally use my Stream Deck thing to like switch stuff around. Like that's how you see me like changing screens while I'm on video and things like that. I think it's going to be a macro pad kind of like this, but specifically designed for codecs. There's really no other information about it other than it's coming on July 15th. It's for Codeex and we just got like a post on X that teases it but doesn't really give us any details about what it is. But I'm a sucker for gadgets. I mean I think if you look around my office that's like pretty obvious that I buy all of the gadgets. So, you know, I'm probably going to get that one too. And since I'm a fan of Codeex myself, it's my main sort of go-to coding platform right now. I'm sure it'll probably be pretty useful. We'll see. I have no idea.
Anyway, that's what I got for you today. There is so much happening in the AI world every single week. It is literally a full-time job to try to keep up with the AI news. But that's what this channel's for. I drink from the fire hose all week. I consume it all. I get overwhelmed trying to keep up with everything that's going on so that I can make one video every Friday that breaks down everything you need to know about what's happening in the AI world. I've also been doing more tutorials and exploration of actual good use cases for some of these tools in between my Friday news videos. So, if those two types of videos interest you, maybe consider liking this one and subscribing to this channel and I'll make sure more stuff like that shows up in your feed. I'm like 25,000 people away from a million and I'm really hoping to get there within the next couple months. So, you could definitely help me get there by just pressing that one little button. That'd be really cool. Sorry, I'll stop begging now. But anyway, my real goal here is to try to save you from overwhelm and also show you some practical real-world use cases of a lot of this stuff that's coming out. I do record these videos on Thursdays and then I go and publish them on Fridays. So, if any new news came out like late Thursday evening or on Friday, well, they'll typically make the following week's video. But again, that's what I got for you. Hopefully, you learned something. Hopefully, you found this helpful. I really appreciate you. Thanks for nerding out with me. And hopefully I'll see you in the next one. Bye-bye. Woohoo. What is up everybody? And we're live. What a way to start the