AI News: Dots, GPT-6.1 Sol, Sonnet 5.5, Gemini 4, and everything you need to know transcript

It's been another busy week between OpenAI's dev day, which I was actually in attendance at, a new model launch out of Enthropic, a new model launchish out of Google, and a handful of other announcements. Yeah, there's a lot to talk about. I'm done wasting your time. Let's go ahead and jump right into it. Let's start with recapping all the announcements they made at Devday because there was quite a bit. Now, the biggest announcement they made was probably about their new AI assistant agent thing called DOT.

They describe it as remarkably capable, always on agents built to handle everything. They also put out a dedicated blog post on this one and essentially what it is is a tool where you tell it what you need and there's no model selector. You don't decide whether to use codecs or work or chat GPT, just chat mode. it goes and decides what to use on your behalf. Kind of runs behind the scenes and then gives you a response.

You can type to it just like you would chat GPT or you can actually call it like you're making a phone call with it. Now, the other thing I will say about Dots that I found pretty cool was that it's very proactive. And what I mean by that is that when an email comes in that it thinks is important, it will actually alert you and say, "Hey, I just saw this email come across. I think you should respond quickly. Do you want me to write a draft for you?" or if it's connected to your Slack.

It will say, "Hey, I just saw this Slack message come across. I think it looks pretty urgent. Would you like me to craft up a reply for you?" Full disclosure, I did get early access to DOTS. Played with it for about a week before the dev day. And that was the thing that I noticed the most about it was how often it was messaging me to tell me to check in on things that might have popped in in my email or Slack or things like that.

So, when you open the chat GPT app, up under new chat, you should see your dot. I literally just named mine dot because I'm not creative enough to think of a new name. And I picked my little wolf character that I made back when they introduced the pets thing as the little like avatar for it. And you can see as soon as I got access to it, it basically went to work for me automatically telling me before I asked anything, here's some things I can take off your plate for you.

RSVP something for you, help respond to an email, reply to another email. So I told it what to do. It actually wrote and sent an email. I then asked it to look through my inbox and said, "Are there emails we can ignore or archive? Are there emails you can help me respond to right now?" And although I probably need to blur most of this out, it went and found a handful of emails that it could just archive that I don't need to reply to.

And then it found a bunch of emails that it could write the reply for me and it drafted up a reply. I specifically told it, "Don't send any emails on my behalf. Only write the drafts and I'll review them before sending." And it went and wrote up a bunch of drafts for me. But here we can see where it starts to get proactive. On Monday, before I ever sent it any sort of messages, it says two heads up for Monday. It pointed out an email and a meeting that I was invited to for Monday.

Then on Tuesday, it actually found something in my Slack that it wanted to remind me about. Later on, I had a flight that was getting delayed. It alerted me about the delay before I even saw an email myself that it was delayed. Then on the actual day of dev day, it actually sent me this thing that it called my personal scratchpad. And when I opened this up, it actually gave me a whole breakdown of my agenda for the day for while I was at dev day.

It collected all the various emails and communications and Slack messages and everything I've kind of gone back and forth with throughout the week about what I would need to know for dev day and it put it all into a document for me. So I just had one place to look to make sure I knew everything that was going on. So all of that was really cool. I really like how proactive it was. Now, the other thing that I really appreciate about it was all of the things that I've already connected with ChatGpt.

So, things like Notion, To-Doist, Slack, GitHub, Granola, Google Calendar, Google Drive, Gmail, my iMes, all of this stuff is already connected to my chat GPT app. And so, that was all the stuff that Dot immediately had access to. So, I was able to see all of my communications across every platform I communicate on basically, and help pull all that information together. So the proactiveness and just like the ease of onboarding because I was already connected to all of these tools made just starting to use it super super simple.

Now having said all of that, there is some downsides to do. The biggest one being that DOTS is available to pro and business premium users in eligible markets. So it's not available in every market and you need to be on a pro or business plan if you need a memory refresh on those plans. Pro plans start at $100. So, the cheapest plan that you can use this new DOTS feature on cost you $100. They also reactivated their $200 a month plan, which they had recently deactivated.

And they added a new $500 a month plan. The $200 a month plan. They actually cut the amount of usage you get with it. So, you don't get as much usage as you used to get for $200. And now if you want to get kind of what you used to have on the $200 plan, they want you on the $500 plan. Now, they still do have their $8 a month and $20 a month plans, but again, to be able to use DOT, this always on assistant agent thing, you need to be on the $100 a month plan.

And I got to be honest, it's kind of a tough cell when Muse just rolled out, which is sort of Meta's version of an always on assistant. And this one's totally free to use. like you just need to have a meta account and you could go and use Muse. Now Muse, to be fair, doesn't have as many connectors. There's still a lot of things that the OpenAI version will connect to that Muse doesn't connect to yet. It doesn't have a granola connector yet, which is one that I do use.

It also doesn't connect directly to your iMessage, but quite honestly, like almost everything else you can think of, it has a connector for now with more coming. So you have dots and you have Meta's Muse and you have Grockbot and all of them kind of do that similar like always on assistant thing. All of them can sort of be proactive for you. All of them could sort of connect to a lot of tools for you but it seems like the new chat GPT version is like the most expensive option.

I will tell you it is good. Is it good enough for most people to use that over a free version of Muse? I don't know. That's a tough sell to me. But it definitely has much better models under the hood than what you're going to get with Muse. Like, you're getting the best models out of OpenAI under the hood. If you tell it to go build you an app, it will go and use codecs and the best model it thinks it needs to use to build that app for you.

If you need it to connect the dots on a whole bunch of different items for you, it'll do that. I had it sift through archives of old emails and try to put together a narrative of some conversations I've had over the years with a specific company and it went and found all the emails and helped me arrange all of these conversations that I've had. So I have like a paper trail of conversations I had with a specific company.

So it is really really good and probably smarter and way more capable than Muse but again at that higher price point. Also, as a quick interesting side note, the website.com is actually owned by Elon Musk and if you go to dot.com, it actually redirects you to Grockbot. I don't know if that was Elon trolling and he knew that that was what they were going to call it and he got ahead of it or if that's just purely coincidental, but I'm leaning towards probably the first one knowing Elon.

Now, probably the second most notable item to come out of Devday was the launch of GPT 6.1 Soul. It's a much less expensive model that the way they described it is almost as good as Astra. Looking at the dedicated blog post for this one here, we can see the pricing via the API is $2 per million token input, $10 per million token output versus Astra, which is $10 input, $50 output. And when it comes to coding, yeah, as we can see, it is pretty much as good, if not slightly better on Deep Suite than GPT6 Astra.

Nowhere near what we're seeing with the anthropic Opus 5.5, but you know, nearly as good as Astra. We can see here on Automation Bench that GPT6 Astra is a bit smarter, but also a bit more expensive on a cost per task basis than GPT 6.1 Soul here. And pretty much all of the benchmarks show similar information. Astra is smarter but more expensive. 6.1 Soul is almost as good but less expensive. To use 6.1 Soul, you do need to be on a paid plan plus pro, business, enterprise, and edu users.

And it's not available in chat yet. You could use it in work mode and codeex, but not specifically just chat. If we take a peek at artificial analysis here, the benchmark that sort of combines a bunch of other benchmarks and tries to give you the sort of objectively smartest model, we can see that 6.1 soul falls down here with a score of 52, which is only one point below GPT6 Astra, but a whole six points below Opus 5.5 here.

And on a cost per task basis, it's a pretty economical model falling at about 72 cents per task where Opus 5.5 and Sonet 5.5 are, you know, way up here at $6 almost $8 and Astra right here at $326. Moving on, another really cool thing they announced was this ultraast mode, which I believe uses the Cerebrus chips. Now, this one's coming soon, but it generates 300 tokens per second, which is about eight times faster than just not the ultra fast mode.

However, Sam Alman did say it's going to come at 6x the price. So, if you want super super fast, we're going to be able to get super super fast AI use. Uh, but yeah, it comes at quite a bit of an extra cost as well. Also, you do have to be on the $500 a month plan to access the ultra fast mode. They also announced private intelligence. With this, you could basically use their models without giving OpenAI personal access to the underlying content.

So, you get confidential computing without them, you know, training on your data or anything like that. They also announced codecs in the cloud. So, you can give it a prompt to go code something for you and it will actually code it in the cloud. So, you can, you know, close your laptop and walk away and it'll keep on going. You can pick up where you left off from like a different computer, things like that. Everything just kind of keeps going in the cloud even if you like turn off your computer.

Let's see a refresh codec cli code review codec security cloud. I'm not going to go into too much depth on this stuff. But then they also announced the decisions API which is interesting because last week I talked about Jev which Jev doesn't output text. It sort of makes decisions. It looks at something and basically selects from a set amount of possible outcomes or true or false or a rating. Well, the decisions API seems to be OpenAI's answer to that. decisions API enables real-time decision-making by focusing Luna's intelligence on a specific set of userdefined questions with finite predefined answers.

So basically exactly what Jev does. If you don't remember what Jev does, check out last Friday's news video cuz I break down Jev in that one. But the decisions API basically does that. You give it information. It looks at that information and makes a decision instead of outputting a ton of text. it just outputs from the predefined options that it has available to it. The agents API now supports computer use. They talked about plug-in extensions so that developers can build on top of chat GPT.

Apparently, they've made it easier to create the plugins and submit them. Now, this is one that I'm actually still a little confused by. They announced ChatGpt spaces as well as pages, and they feel very similar. So, ChatGpt spaces, a new home for your team to collaborate with AI to get work done. Create a dedicated space where teammates, chat GPT, and your DOT can build on shared knowledge. ChatGpt can keep your space organized based on instructions you provide, so you can quickly find what you need and pick up where you left off.

Available to Pro, business, and enterprise plan. And then we have pages. Pages are a new type of document built for human and agent collaboration. Anything you can do in chat GPT, you can do on a page. write, research, generate charts, create images, or visualize information. So to me, pages almost feels like notion but built straight inside of chat GPT where chatg spaces almost feel similar to what projects already are, but more collaborative.

That's sort of how I'm interpreting it, but I could be a little off on that. They have collaborative slides, create teams and share tasks, chat GBT and Slack and Teams, a meetings plugin, which seems like basically a new competitor to Granola. So like it just seems like it does exactly what Granola already does. Takes notes for you and saves personalized summaries with action items in chat GPT space. Audio is deleted once your notes are ready and can't be accessed or replayed.

This one was interesting as well. Sign in with chat GPT. So if you are developing something, you can let people log in with their chat GPT account now. And if the thing that you built uses AI, it'll use that person's API key that they logged in with. So it won't charge you and your API key, it'll charge theirs instead, which seems kind of cool if you're building some sort of like AI tool. Again, they announced the new Pro 500 tier, the $500 a month tier, which also gets you access to the Ultra Fast and OpenAI marketplace, which this to me feels like the third or fourth time they've attempted to promote a marketplace like this, and I'm still a little confused about what the difference between this marketplace is and the last couple iterations.

But if you build tools, it seems like you could build stuff and put it inside the OpenAI marketplace. So, again, a lot a lot of announcements. Also, fun fact, if you watch the replay of the keynote and you fast forward to like 51 minutes or so, you can actually spot me in the crowd. Proof that I was there. Another cool update comes from Optimizely. I talked about their new virtual teammates in a previous video, but I've been able to play around with their preconfigured SEO and AI search analyst teammate, and it's pretty cool.

I just connected it to my futuretools.io io site, set up the virtual teammate, and then I asked it, "Audit future tools for SEO and AI search visibility and tell me what to fix first." It can research keywords, find content gaps, and check how ready my site is to show up in AI answers. That's the AEO part of it that everyone is focusing on these days. It stands for answer engine optimization because so many people are getting their recommendations from AI instead of traditional search.

Now, this virtual teammate helps me with all of that, and it can continue to monitor the site and proactively collaborate with me on prioritized action items, so I've got a running list of what needs attention without manually digging through dashboards and figuring it out all on my own. If you want to check it out or any of their other virtual teammates, head to the link in the description box. And thanks so much to Optimizely for supporting my channel and sponsoring this portion of today's video.

But the sort of flurry of announcements that we got out of dev day was not the only news we got this week. Enthropic had a big launch this week. They announced Claude Sonnet 5.5, which is their new model that's cheaper than Opus 5.5, but you know, almost as good, kind of like 6.1 to Astra. If we look at their benchmarks down here, you can see on Terminal Bench, it's now their top scorer. But on pretty much all the other benchmarks, Opus 5.5 beats it out.

But pricing is the big difference. It's about half the cost of Opus 5.5. $4 per million input tokens down to $2. And then on output tokens, you've got $20 down to 10. So, it's quite a bit less expensive to use. Looking at their chart here, you can see the further to the right we get, the higher the cost to use it. Orange is Opus 5.5, blue is set 5.5. Interestingly, Sonnet is actually more expensive to use at max than Opus is to use at max, but it's also a little bit smarter at max.

Now, when we go back up to these benchmarks and we look at the system card on how they got these benchmarks, I actually get a little bit confused because it does say, unless otherwise noted, all results use the following standard configuration, adaptive thinking at max effort. So when we jump back to these benchmarks, these are all at max effort unless otherwise stated. And when we look at their charts down here, max effort is actually more expensive to use.

Right? So if I click through all of these, we can see max is the most expensive setting. And in most cases, max is more expensive than using opus. So these benchmarks are showing Sonnet at max effort getting these benchmarks which underperform Opus 5.5 but theoretically in most cases is actually costing us more than Opus 5.5. I don't know if I'm making sense here but like why are they benchmarking this at max if max is actually more expensive than using Opus 5.5 and showing us the benchmarks based on max effort.

And where this is even further confirmed, if I jump over to artificial analysis here, we can see that Sonnet 5.5 is 56. It's it comes in second place behind Opus 5.5. But if we jump down to our cost per task here, Sonnet 5.5 on a cost per task basis averages $7.62 where Opus 5.5 averages $5.98. So on average, Opus is cheaper to use than Sonnet. This is also max effort, which is the same effort level that they used to get these benchmarks.

So, I'm just very confused. Like, Sonnet 5.5 seems like a step down, but to get these benchmarks was more expensive than Opus 5.5. When we look at output tokens per task, Sonnet 5.5 is the most token hungry model out of all of them, using on average 194,000 tokens at max effort level. Now, I pointed this out on Twitter and Edwin here who works for Enthropic said, "Do not use Sonnet with max effort. At that point, you should probably be using Opus." Okay, I get it.

Max effort is token hungry and we should probably not be using max effort. If we're at that level, go use Opus instead. But then their benchmarks again, what they're showing when they're showing this is how capable our model is, they're showing it using max effort. If you don't think we should be using max effort, maybe like benchmark it based on the effort level you think we should be using. Okay, I'm I'm done ranting about this.

We got a new sonnet model. It's almost as good as opus. It's more expensive if you use max effort. Less expensive if you use a lower effort level, but if you're using a lower effort level, then these benchmarks aren't really relevant to you because they only were able to get these benchmarks using a max effort. So yeah, that's Sonet 5.5. I of course did beauty bench both of these. Here's Claude Sonnet 5.5's result. Used 29,000 tokens and cost 28 cents.

GPT 6.1 Soul generated this version 35,000 tokens and cost 34. And our LLM as a judge actually put 6.1 Soul as the new leader and ranked Sonnet 5.5 all the way down here at number 13. Some other quick announcements out of Claude. You can now customize Claude code with mods. So if you want to like actually modify the claude code platform, you can actually prompt it to essentially mod itself. We can see on Twitter here they posted about it.

You can now mod cloud code, change how it behaves, customize the UI, swap in your own features. So now you can basically prompt inside of Claude Code to essentially have it tweak Claude codes functionality, which is kind of cool. We also got a model out of Google DeepMind this week kind of. They made an announcement post for their new Gemini 4 Argon model, which if you look at the benchmarks on this one, pretty much makes it the new state-of-the-art model above Opus 5.5, above GPT6 Astra.

Before this model was announced, Opus 5.5 pretty much owned every single benchmark. According to their new article, this new Gemini 4 Argon is pretty much state-of-the-art across the board. However, this blog post is announcing their new Frontier model, Gemini 4 Argon, but it's only rolling out to a set of trusted cyber defenders through their Fair Wind program. So, they announced this new model and they said it's here, but most people can't have it yet.

The pricing of it is pretty dang reasonable. It'll launch it at an introductory price of $2 per million input tokens and $10 per million output tokens, which is great pricing for how good this model is. And something that's really interesting as well is that they say they're significantly expanding the model's output token limit to an industry-leading 1 million tokens up from a previous of 64,000 tokens. This is 1 million tokens of output which is huge.

Before it's typically like 1 million combined tokens where most of that is input tokens and like 64,000 of it being output tokens. Now it can output over a million tokens. That is just crazy. It can output over 750,000 words essentially. That is just wild. But again, it's a model we don't have access to. But when it comes to knowledge work, new state-of-the-art, when it comes to coding on Deep Suite, this would make it the state-of-the-art.

And yeah, pretty much across the board, it would be the best model out there. But again, rolling out soon. So, we don't have it yet. And when we do get it, it's going to go to API customers and Google AI ultra subscribers first. So, their highest tier plan. Interestingly, Artificial Analysis apparently got access to this and was able to test it and it actually scores tied with GPT6 Astra on Artificial Analysis here despite the fact that all of those other benchmarks on their own website put it as pretty much the top dog, best of the best model.

But according to Artificial Analysis, which you know combines a bunch of benchmarks, it's about as good as Astra. I'm sure once it's more widely available, it'll actually get benchmarked even more on more tests and maybe this number will update. I don't know. We also got a new image model out of Ideogram with ideoggram 4.5. And apparently this is the most precise editing model. It appears to be really, really accurate at taking other images and making any edit you want, no matter how big or small, to the image.

You can actually try this one out over at idiogram.ai. So, I'll start with an image that's already in here that somebody else generated. And I'll drag this image into the reference image here. And then I'll say put him at Petco Park in San Diego. We're on ideoggram 4.5. Let's go ahead and generate that edit. And it gave me four variations where he's definitely at a ballpark, but the original person looks exactly the same.

So, it did a good job with that. When it comes to new image models, I find it harder and harder to be impressed by them because I feel like they're all pretty good now. But IOG 4.5 is apparently the leader in editing. So, if you need to make edits to an image, there's a new one out there for you to try. All right, we've covered a lot already between Dev Day and Anthropics announcement. It was a lot of news this week, but there's a few more things I want to share, but we don't need to go into too much depth.

So, let's just break them down in a quick rapid fire. Pew. Starting with yet another decision model. Last week we talked about Jev. Earlier on we talked about the OpenAI decision model. Well, there's also a model from this company Strand which is called Strand Decider 2B which is a small opensource decision model. Now if you don't know who Strands is, don't worry. I didn't either. But they're actually just Amazon. It's like Amazon's agentic experimental AI division.

I don't know if there's great benchmarks really yet for these decision models. I don't quite know how to read this. This just says right and down is better. So yeah, I guess that's good. I I don't know how to read these benchmarks yet, but apparently it's decent. But yeah, we have another decision model out of Amazon's division called Strand. SpaceX AI announced team bots, which appears to be like Grockbot, but you know, available for teams.

You build a team bot around a role or workflow your team shares. Everyone gets access to the same bot and its expertise while still being able to work with it individually. So all the team bots get the same context, plugins, credentials, and memories. And then team members could then access that bot. So if you're a big fan of Grockbot and you want to work with your Grock bots collaboratively, now it seems like these team bots are designed for that.

We got two new voice models this week. 11 V4 from 11 Labs, which is apparently their most emotive model. This new 11 Labs model seems to be pretty good. You can tell it how you want it to sound. So like excited and happy sounds like this. >> Big news, everyone. Our summer menu drops this Friday. And yes, the mango sticky rice is finally back. >> If you wanted to sound sleepy and drowsy, five more minutes. I promise I'll get up right after this dream finishes. >> And it's got, you know, other languages.

And apparently it's better at actually cloning your voice as well. So like here's the real voice. >> Hey, thanks so much for holding. Um, so I've got your account pulled up now. >> And then here is their version. >> Hey, thanks so much for holding. Um, so I've got your account pulled up now. >> And I mean, yeah, it's pretty dang close. But Microsoft also released their own new voice model called MAI Voice 2.1 this week, which also has realistic expression and multiple languages.

Here's an example. >> Here's the clever part. Plants capture sunlight and turn it into food. >> Here's another example. >> Here's the clever part. Plants capture sunlight and turn it into food. >> I mean, comparing them side by side, I would say maybe 11 Labs has the edge, but you know, you have another option. Now, also from Microsoft AI, they have a new streaming transcription model, which basically means it'll transcribe really, really fast.

So, you can be doing like a live stream or something like that and it could transcribe straight to the screen like in practically real time. And finally, a bunch of AI leaders met with the White House this week and they all agreed to rebrand AI as super intelligence. Now whether that actually happens in practice, we're still yet to see. But we can see it was signed by Sundar, Daario Zuckerberg, Greg Brockman, Elon Musk, and Jensen Hong.

This actually happened the same day as DevDay. So Greg Brockman was at the White House while Sam Alman was at DevDay, and it was also signed by Donald Trump, who of course is president of the United States. Other than agreeing to call it super intelligence, though, they also agreed that we're all going to build the technology safely. and they agreed on four layers. Implement robust internal controls to monitor the capabilities and alignment of its models during training and deployment.

Empower an internal team to ensure all of the controls, monitoring, and detection are operating as intended. Partner with an independent external auditor or evaluator and designate an independent committee of the board of directors to oversee and receive reports from the teams operating the controls, etc. And again, they all signed it, but it's not anything that's like super binding or enforcable either. Anyway, that's what I got for you today.

I had a blast at OpenAI dev day and then previously Meta Connect last week. It is always an interesting thing where you go to these events and you meet all these other creators and you meet the people building the tools and then you come home and you know want to report on everything that you learned, but you're in this like afterlow of sort of excitement from hanging out with all these people. But then you need to come to reality that when you're in the event, everybody's sort of got this buzz going and excitement and everything seems really cool in the moment.

And then you get home and reality sets in and you start testing it and you're like, is this as good as you know what it was made out to be at the event? And then you got to go and make content and tell people what you really think. And it's just always like this really weird thing after events because my responsibility is to make videos that are the best possible video for the audience to understand what's going on. But while I'm at those events, you know, they're giving me free GameBoys and, you know, they're providing us with meals and hotels and we're hanging out with all these other creators and we've got this like fun buzz going on where we're all excited and it does make it a little bit more difficult to come home and be like, "That wasn't that cool of an announcement actually." But again, I'm always going to do my best to try to, you know, do right by the people who watch my videos and call it how I see it.

And sometimes that might mean that in the future I don't get invited to as many events, but you know, it is what it is. My goal is to make sure that people watch these videos and understand what's going on and get my actual real take and not the take that some of the companies probably wish that I gave. But again, that's what I got for you. It is my goal with this channel to drink from the fire hose and be overwhelmed all week and learn everything that's going on in the world of AI and then turn around and try to condense it and distill it and try to make it as understandable as possible for you so you don't have to feel overwhelmed and like you're drinking from the fire hose.

I do record these on Thursdays and publish them on Fridays. So, if any new news came out late Thursday evening or on Friday, it'll probably end up in next week's video. But if that's something that sounds interesting to you and whether you love AI or you hate AI or you're indifferent to AI, I'm going to do my best to try to help you keep up with it and give my just real raw thoughts. I don't script any of these videos.

I just sort of blurt out whatever is on my mind as I'm talking about it. And what comes out is honestly my raw unfiltered thoughts on everything I saw. And I often also change my mind later on too. I might say something's really cool and then two weeks later after using it even more go, you know what? That actually isn't as great as I thought it was when I first played with it. So, you know, stick around to see how my opinions change.

Again, really, really appreciate you hanging out with me, nerding out with me. Lots going on. Luckily, I'm done with traveling for a while. I'm not going to any more events for a while. I'm kind of saying no. I want to just like heads down, learn as much as I can, play with the stuff that's all been launched recently, and make even more videos to help you understand what's going on and show you cool use cases. So, again, that's something of interest.

Like this video, subscribe to this channel. I'll make sure more stuff like that shows up in your YouTube feed. Thanks again and hopefully I'll see you in the next one.