Claude Opus 5.5 Didn’t Need to Go This Hard transcript

Another day, another new best model in the world just came out. And of course, on the day where we get not one but two state-of-the-art models out of the big foundation labs, I'm out in Palo Alto for Meta Connect. So, doing a quick review of what just came out from my hotel room. So earlier this morning, we got a new model out of Anthropic, and it appears to be the new best of the best, better than Fable 5.1, but cheaper and faster.

And this model is Claude Opus 5.5 out of Enthropic. They claim on their website it now performs at the level of Claude 5.1 on most work and costs 40% less to run than Opus 5. So, better than 5.1, cheaper than Opus 5, the, you know, tier below Fable. So, let's scroll down to the benchmarks here and take a quick peek. Again, I don't put as much weight on benchmarks as I used to. I put much more weight on what we can see and test and feel and what other people are actually doing with this.

But, let's take a peek real quick nonetheless. So, as far as agentic coding goes, 66.4% that makes it state-of-the-art across the board. Agentic Coding on Frontier Code 54.4% state-of-the-art across the board. Better than Fable 5.1, which Fable is supposed to be their top top tier, right? That's supposed to be the best of the best in their model lineups. And now Opus, the tier below it, is actually performing better in most areas.

Agentic Coding on Cursor Bench 57.8% makes that state-of-the-art. GDP val for knowledge work 1846 state-of-the-art across the board. Automation bench Astra still has them beat there. Humanity's last exam new state-of-the-art terminal bench science still give the lead to Astra here. Computer use state-of-the-art visual chart recognition state-of-the-art. So, Opus 5.5 beats Fable 5.1 and Opus 5 pretty much across the board and it beats out GPT6 Astra in most areas.

And here's what people probably mostly care about is not only does it do most of that stuff better, but it does it for cheaper. So, we can see here input tokens went to $4, down from $5. Output tokens went to $20 down from $25. And that's based on the Opus 5 pricing. Now again, remember they're saying that this is pretty much better than Fable 5.1 for most things. So if we look at Fable pricing here, Fable 5.1 is $10 per million input tokens, $50 per million output tokens.

So $10 for Fable 5.1 down to $4 for Opus 5.5, $50 for Fable 5.1 down to $20 for Opus 5.5. So, if you're comparing like model capability to price, this just got a lot cheaper for the capability. When it comes to Terminal Bench here for Agentic Coding, we can see cost per score on Terminal Bench. Along the bottom, we've got the price. Along the left side, we've got the score. So, if you're using extra high here, it actually appears to be slightly better score than on max and cheaper.

And then we can see Fable 5.1 is this green line here, which not nearly as smart for way more expensive. As far as availability goes, it's now available pretty much everywhere. So, you should be able to use it in Claude right now. Another benchmark I like to look at is artificial analysis. They're trying to rank like the objectively smartest model based on an aggregation of a whole bunch of benchmarks. And if we look at this, we can see Claude Opus 5.5, now takes the lead with a score of 58.

Fable 5.1 was the previous best at 53, tied with GPT6 Astra, also at 53. So, new smartest model in the world at way cheaper than the previous smartest models in the world. Scrolling down here to cost per task, we can also see that it is cheaper than Fable 5.1, but not by a ton. So we can see Fable 5.1 cost about $763 per task where Opus 5.5 in max mode cost about $5.98. So still pricey and the reason being yes it got cheaper per token but if we look at the output tokens per task Claude Opus 5.5 now uses way more tokens.

So if we look at Fable 5.1 it used to use 78,000 tokens per task. Opus 5.5 uses 119,000 tokens per task. So, it's using a lot more tokens at a cheaper per token price. I mean, at the end of the day, you're still getting cheaper per task, and that's what really matters. But the reason it's not such a big drop off here is because it does use so many more tokens. Okay, now that we've got that out of the way, let's take a peek at what some people are actually building with Opus 5.5 because the demos have been really impressive.

So Drew here made this animation using Opus 5.5. This whole thing was created with JavaScript with Opus 5.5 and it looks impressive. This is an amazing animation. Like I did not think we would get to a point where LLMs could generate animations like this without needing to go into something like After Effects and generate them. So super impressed with what we're seeing on the screen right now. Here's another example of an animation from Kevin.

No or go. I'm not sure how to pronounce that, but Opus 5.5 drew every frame of this animation in JavaScript. Everyone in town sends Claude their request, but one girl sends a question instead. What do you love? And this looks like a, you know, children's drawing sort of animation. And once again, ultra ultra impressive animation. Like I didn't think we would get to this point. Angel here with another demo. They did what they call the Game Boy test.

I guess they had it generate a Game Boy and then actually use the Game Boy. And I'm going to start to be really really repetitive here because everything I'm seeing come out of Opus 5.5 is insanely impressive. This Game Boy little emulator here really impressive. Edwin here, who full disclosure actually works at Enthropic shared this demo of a game inspired by anti- Cathedral mechanism. I don't know if I'm saying that right, but once again, like crazy crazy impressive demos.

Like this animation was all made using just Claude. Like this this is just it's it's blowing my mind. It is. It really is blowing my mind that we're getting this kind of graphics right now. Here's another animation that was generated in Blender. Alex Albert, who I believe also works at Anthropic, generated this animation and did the whole thing using Blender. They asked it to do a claimation look, render the video, and then check the frames before sending it.

Now, I'm not sure if this actually used the MCP or if it used computer use, but either way, we're getting really, really good results out of Blender using Opus as well. Tac here showed off his firsterson shooter doodle game, which also looks really, really impressive. Hakam here showed off his version of Snake that was made with Opus 5.5. I just like the look of it sort of dragging through the sand. And you can see that over time the trail gets thinner and thinner.

So, I don't know, just like really, really cool details. Again, hard to be impressed by a snake game, but uh this one's pretty decent. Alex from Forward Future made a version of Dark Souls, and I mean, holy crap. Like, this looks insane to me. Like if somebody sent this to me a year ago and said I generated this with AI, I would have assumed they were trolling me and that this was just a clip from an actual game that they were pretending was made by AI.

But you know, I know the Forward Future guys, that's Matt Berman's team. They're not faking this. They made this with Opus. This is insane. Here's a flight simulator that Alex from Forward Future also made. Let me fast forward a little bit. You can get sort of a firstperson view out of the cockpit. And again, it's hard to not be impressed by what people are making with this right now. Alex also got it to make a version of Mario Maker, which looks pretty decent as well.

I mean, yeah, I I know I'm sounding like a broken record just saying how impressed I am over and over again. I think looking at games and animations are two of the best things to look at with these models to show how far we've come. Because when I look at sort of games and animations that these things were doing just a year ago, we were nowhere near this. Like they weren't even making animations at all. Games were like decent, but they were fairly archaic looking.

Now we're getting stuff that looks like what you'd get out of a studio today. Just super impressive. But here's the thing. Opus 5.5 wasn't the only model we got today. About two hours after Opus 5.5 was announced, OpenAI jumped into the party and said, "Hey, we got a new model today, too." And they launched GPT6 Soul and Luna. Now, I'm going to be totally real with you. These new models aren't quite as impressive as what we just saw out of Opus 5.5.

Now, they're good. They're less expensive. They're faster, but they're not quite as state-of-the-art as what we just got out of Opus 5.5. Now, the article that they put out is a little bit light on the benchmark side. We can see their pricing here for GPT6 Soul versus 5.6 Soul. They dropped the price on this new model from $4 per million input tokens to $2 and their output price from $20 per million down to $10. So, they slashed the price in half on GPT6 Soul from 5.6 Soul.

Luna quite a bit less expensive as well. They slashed the input price in half from 20 cents to 10 cents and the output price from a$120 down to 50 cent. So actually more than 50 cents cheaper on output. Again, not a ton of benchmark data here. They did cherrypick a few benchmarks. We've got Automation Bench here where they show GPT6 Soul scoring not quite as good as Astra, but quite a bit cheaper cost per task. So on extra high here, it scores about as good as Astra medium but at a cheaper cost.

When we look at agents last exam, you can see the blue stars are Astra, the yellow stars are GPT6 stole. So, as far as score goes, their max model in GPT6 does better than Astra's low model at about I don't know roughly the same cost, if not a little bit cheaper. On Deep Suite, the coding benchmark, it does decent as well. So, it scores better than Astra on low, but at a slightly higher price than Astra on low. So that's why this model feels a little more marginal than the Anthropic announcement because it is a cheaper, better soul model, but it's not as good as Astra.

And if we go look at what Anthropic released, they released a new Opus model that is actually currently better than their previous state-of-the-art Fable model. So, when we're looking at these two launches that came out the same day, you got to give that edge to Enthropic a little bit because they just put out a model that's faster, cheaper, and better than their previous state-of-the-art, where OpenAI put out a model that is cheaper and faster, but not nearly as good as their state-of-the-art.

Now, as far as availability goes, GPT6 Soul and GPT6 Luna are available in ChatGpt Work and Codeex starting today if you're on one of the paid plans, plus pro, business, enterprise, and .edu EU users, free and go users can use GPT6 Luna in the desktop app. Now, it also appears to me that GPT6 Soul and Luna weren't really distributed to a bunch of early access people cuz I couldn't find any demos of people using this yet.

Fable, a lot of people seem to have a little bit early access, got to play with it, and now started showing off their demos. I wasn't somebody that got early access to either of these, so I have no demos to show. Otherwise, I would have probably recorded those back at home and not from my hotel room here at MetaConnect. But when I do scroll X today, almost everything I see on X is all about Opus 5.5 and very, very, very little about GPT 6 Soul.

So, not nearly as many benchmarks, probably because they know it's not going to top the state-of-the-art Opus 5.5 and it's not going to top Astra. Also, not a lot of demos circulating because from what I can tell, not a lot of people got early access. I'm sure some people did, but because Opus 5.5 was so much more impressive today, almost all the demos I'm seeing are out of Opus. Taking a peek back at artificial analysis again, something that's supposed to be like the objective smartest model.

We can look where GPT6 Soul falls, and it falls at a 4.8, making it tied with Muse Spark. But you can see that GPT6 Astra is still at 53 and the Opus 5.5 we got today is 58. So the new anthropic model is a full 10 points higher than GPT6 Soul. And then GPT6 Luna falls all the way down here and appears to be tied with GPT 5.6 Luna. When it comes to cost per task, it is a fairly inexpensive model. GPT6 Soul down here comes out at a$16 per task where if you remember Opus 5.5 is all the way up here at almost six bucks per task.

And as far as output tokens used, it falls all the way down here at 31,000 per task. If you remember, 5.5 max is all the way up here at 119,000 per task. But again, I feel like cost per task is a more important measurement. You just want to know how expensive something's going to be to run. you probably don't care as much how many tokens it's actually using. And as far as cost per task, six soul comes out to about a$16.

Oh, real quick, I forgot to mention I did run both of these models through beauty bench. And you can see Opus 5.5 here, GPT6 Soul here, and then Soul Pro here. Interestingly, if I take a look at my leaderboard, it actually puts GPT6 Soul as the new leader with GPT6 Soul Pro a few points below it. Now, again, take these leaderboard rankings with a grain of salt. Use your eyes to decide what you think is best. This is an LLM as a judge deciding this.

I personally think Astra still looks better than GPT6 Soul to me, but my LLM as a judge put GPT6 Soul as the new leader. If you're wondering where Claude Opus 5.5 lands, it put it down here at number eight. But I don't really pay attention to the leaderboard. I pay attention to my eyes. So, that's what I got for you today. I wanted to quickly break down these two model launches that both happened on the same day. This is going to be an insane week by the time the week is over because again, MetaConnect is happening right now as well, which we're going to get some announcements out of this event.

We've got Made on YouTube happening this week as well, which we'll probably have some announcements probably more for YouTubers. There's also a ton of buzz about Jev, which I will be talking about in another video. And of course, I will do my Friday news breakdown just like I do every week. We will dive more into Opus 5.5 into GPT6 Soul. Whatever news comes out of Meta Connect this week, we'll talk a little bit about Jev and all of the other announcements that end up happening this week.

Those will all be in my Friday news breakdown. By then, there will be much more demos from some of these products. We'll have had a chance to do even more testing with them, and we'll be able to go even deeper. I just wanted to give you the initial impressions, what people are doing with them, let you know that these are out there and available if you're on a paid plan for either of these companies. You can be using these models right now.

If I was to sum this all up in like a couple sentences, Anthropics Opus 5.5 came out today. It's way cheaper, way faster, and apparently way smarter and better than Fable 5.1, their previous state-of-the-art model. So, their next version of Fable is going to be even better than this model, which is just going to be mind-blowing. And then, because OpenAI needs to also stay in the news cycle and not be overshadowed by anthropic, OpenAI released GPT6 Soul today, a couple hours later, which if we're being honest, feels like a fairly marginal improvement because it's not as good as Astra and it's slightly better than the previous generation of Soul models.

However, yes, it is faster and cheaper, but you're probably not going to notice a huge difference using this model over something like Astra, other than it'll be less expensive to use than Astra if you're using the API. So, fairly marginal update unless you're a developer who's looking for a cheaper GPT model, but the big news here is definitely Opus 5.5. It is really damn impressive. Again, that's what I got for you today.

Every week, I keep up with all the AI news. I drink from the fire hose. I overwhelm myself with all the news and then every Friday I make a breakdown video of here's all the news you missed this week. So if you want to just keep looped in with just one video a week, that Friday video is the video to check out. There will be one coming this Friday that will recap all of this news plus everything else that happened this week.

So if you haven't liked this video and subscribed to this channel, you should probably do that cuz that'll make sure those videos show up for you and uh you'll be totally tapped in. So, thanks for hanging out with me here in my hotel room in Palo Alto. I'm about to head off to MetaConnect right now. Just wanted to get this video out quickly for you to share some of my quick thoughts on it. But, uh, again, thanks for nerding out with me.

Hopefully, I'll see you in the next one. Bye-bye.