Claude Opus 5.5 Is Here And It's INSANE (This Feels Like AGI) transcript

Yeah, Anthropic has done something serious with this model. When GPT6 came out, I thought it felt like AGI, but this just feels different. This really does because it's not only the world's smartest model now, Opus 5.5, but it's also a much faster model and a much cheaper model than Fable 5.1. And right now in the background, and I'll show you guys the output later in the video, I'm running a test where I built a trading engine, which is something I like to do in my spare time, using Claude Fable 5.1, and it's been running for 20 minutes.

Opus 5.5, oneshotted the task in 12 minutes, and only cost a few dollars. I think Fable's probably going to run me like $10. Let's see. It's not even done yet. So, the experiments I've run so far since getting access to this model are suggesting that not only is this a model that's just as smart and actually technically on the intelligence index smarter than Fable 5.1, but it's also a model that's much faster and cheaper, which is pretty scary when you think about it.

And I think the thing that is making these new models feel so special, so I'm specifically referencing GBT6 Astra and Opus 5.5 here, is the fact that I feel like the major labs, Anthropic especially, have solved for RSI or at least very close to it, which is recursive self-improvement, enabling them to use their smarter models, for example, Anthropic has mythos 2 to train and help the other models self-improve. So essentially, you have the smartest models training their next successor.

And it's just a pattern that's going on and on and on. As Alex Finn points out here, and I'll get into some Opus 5.5 examples shortly, it feels like there's some sort of magic behind it that's hard to describe. It's the first model I've used that improved and became near Frontier on basically every metric while also getting faster and cheaper. Because keep in mind, Opus 5.5 is not a frontier model. It's not a fable model, but they are achieving fable-like intelligence at a fraction of the cost.

Honestly, I think we have to start preparing mentally for what's going to happen when they launch their next frontier model because it's very likely that model is being used to train this model. It's also very likely that they are taking their time with releasing it, gatekeeping it, because a it introduces massive security risk releasing it to the public. So it probably takes some more time to implement specific safeguards that are needed to make it public whilst also enabling them to train other models like they don't need to you know Apple operated in the same way.

They would basically be like two iPhones ahead in terms of design features and every new release was really technology that they had for like a prior year or a prior 2 years. So Steve Jobs was known for this. he would plan multiple years ahead and Apple likely already has their plan for a 2029 iPhone or a 2028 iPhone now. It's the same concept here. They don't release the best models that they have. They release models that are trained pretty much 6 months ago and they keep the advantage by using these new models to train their other models and then if the competition catches up like GPT for example in throughanthropics lens starts to push them which they definitely have over the last few months.

Even I as a longtime Claude user was starting to switch to GBT for a lot of tasks. Then they just, you know, put the foot on the gas and it released something great. So it still shows that Anthropic can compete and I think we're going to continue to see this tugofwar. But let's go through Opus 5.5 specifically. Opus 5.5 according to Claw is the first model since we called pacing the frontier. As with previous models, it was tested by external evaluators before release. on our most comprehensive alignment test.

It achieves the strongest score to date. Here is the score benchmark. Opus 5.5 is a major step up from Opus 5 leading on aentic coding computer use and knowledge work. You can see it's not only better than Opus 5. For Agentic Coding, it beats 5.1 and it beats Astra. For knowledge work, it beats 5.1 and it beats Astra. For multi-disciplinary reasoning, it beats Fable and it beats Astra. There are some things where it's still behind.

Astra is better in terms of business workflows slightly. Astra is better in terms of scientific research slightly. Visual chart recognition is better than Fable on Opus. Computer use is better on Opus than Fable. So you could see that it's winning across, you know, a variety of categories. And since using it, I could say anecdotally this is something that I am also experiencing. I'm not experiencing a trade-off in intelligence.

It does feel pretty similar to FA Fable 5.1 in terms of intelligence. it it feels slightly better in some instances, but it feels similar. But what is kind of the unlock that I'm finding is just how much faster it is and obviously how much cheaper it is at the end of the month. That's definitely going to make, you know, a major difference to how much I'm spending on credits cuz I think last month I spent like $5,000 on credits.

I've gone AI crazy. I spend way too much. Obviously, I love experimenting with it. I make videos on it, so I guess you could call some of it like business, but yeah, I spend way too much on AI. So, I'm probably going to see the difference at the end of the month, but what I'm already experiencing in the shorter term is how much faster it is, which literally right now I'm running a back testing engine. Fable is still running.

It's not done yet. It's been like 25 minutes. Opus finished like 12 minutes in. So, that just gives you an example. And uh we'll actually compare the results at the end of this video. So, now let's compare Opus 5.5 to some of the other models. You've got Opus 5.5 in orange, you've got Astra in gray, and you've got Opus 5 in yellow. You can see it clears Astra for business workflows by effort level. In Agentic Coding, it scores higher on most effort levels and it doesn't cost as much as Fable 5.1 or Opus 5.

As you can see, they cost more on the X-axis here. In terms of real world knowledge tasks by effort level, Opus 5.5 clears the pack. In terms of agentic terminal coding, Opus 5.5 clears the pack. And something really interesting that they did is they changed the way it interfaces and communicates. I know, you know, I've never really been a fan of the M dash. It's quite an obvious tell of AI. If you're wondering what that is, especially that like long dash, which you can see here, I'll just zoom in on my screen a little bit.

That is very common in AI writing. They've they've actually gotten rid of that. It's only really using full stops and commas now, which is, I think, a more natural way to write for a lot of people. So, the text does seem a bit more human now. The text is a lot shorter and more concise. I mean, look at how much smaller this input is versus Opus 5. I've already been stacking custom repos on top of Claude. Anyway, I did a video on this yesterday.

Uh I've got this thing called caveman mode and then I've got I have ADHD which basically tells Claude to respond in a much shorter fashion. It's been saving me a lot of tokens, but I think the text reduction by default will save people tokens without needing to install these custom repos in the first place, but I'll probably still run them anyway because I prefer working on KBAM mode when I'm uh developing. It's it's basically a repo you can install because it's just so much quicker uh in terms of interfacing with Claude.

So I've had about 10 hours to test the model at the time of recording this video, but obviously there are some developers which have had early access to this. As you can see here, Powell has had access and has been testing it across multiple development applications. He still finds that Astra for his specific use case, which is hunting bugs, came out on top, but Obus 5.5 was now the best model in the anthropic lineup and it came in cheaper. $60 versus Fable 5.1 which is $77.

I still think there's going to be a use case for GPT6 Astra. It's it's a great model. I think if you were using Claude and you were thinking of switching to GPT, I think there's now a reason to maybe persist and use Opus 5.5 for a week and see if you like it without paying for another subscription. But of course, if you're AI crazy like me and you have multiple subscriptions, you'll tend to just siphon specific forms of work depending on what you want to get done. for example, I still think mass agentic coding, bug fixes, um, GPT is still probably going to be the go-to, but I prefer interfacing with Claude for things like brainstorming, idea generation, creative work, uh, design work.

So, I still feel like the models have strengths in different areas. I do also prefer the UI of Claude, just how they've set up the schedule tasks, the uh, the phone app, the projects. I think it's a little bit more user friendly than GPT in my opinion, but that's like a small thing. I guess you know once you're getting into coding and software development then um you know you're probably coding on VS Code anyway or you're probably just using the terminal.

So you just want the best model per task and Opus 5.5 is going to be you know a very good model for a lot of tasks. Now just looking through some visual tests while my tasks are running in the background as well. Uh it definitely passed the Minecraft test. GPT6 Astrogit as well. But I mean, it's crazy to think that just a few months ago, like it could barely make a Minecraft game. Like, they looked terrible. And now, like, I wouldn't say this is one to one, but if this was like a one-shot prompt, this is pretty damn incredible that AI can now do this.

Matt Schumer says, "I forgot how freeing it is to use Opus and not worry about hitting my limits every 5 minutes. I'm so much less stressed this way." Uh, I definitely think, you know, for the average person, this is a lot more affordable. So, it just makes sense to use Opus 5.5 now over Fable. You're going to, you know, save money and you're going to get pretty much the same outputs. And in some cases, it'll probably seem smarter.

I think something that is actually underlooked or underrated in AI is speed. Because everyone talks about, oh, the smartest model, the most capable model, but speed is actually a really important factor because if you are developing and you have to wait an extra 10 minutes per prompt to get an output, it slows you down. And the compounding effect over a multi-week or a multi-month project is literally days of time. And I'm f and I did find that Fable was very slow.

And Astro can be slow sometimes, too. I think it is faster than Fable, but Fable is like painfully slow. Even like prepping some stuff for these videos would take hours. Running a back test on a trading strategy would take hours. So if I'm able to cut that time, not [clears throat] necessarily even in half, but by 20% 30% 40%, the compounding effect is massive if you can get a similar level of intelligence. So I think people always want to talk about oh what the smartest model is.

In reality, the best model for you is something that is usable in your daily life. So if you do even trade off a few percent of intelligence, you know, you just use Fable as a planner or an orchestrator and then you use another model for speed, that's often times the better solution in many cases. If you go on to artificial analysis.ai, there's actually a model recommender. So you can also put in your preferences. You know, maybe you don't value intelligence as high.

You view it as very important but not critical. Maybe you view speed as critical and maybe you view cost as uh a medium level of importance. And that's going to change your recommendations. So, you know, if you have a certain set of criteria, Quinn could actually be a better model or GPT could be a better model or Muse, you know, for aentic work could be a better model depending on what you want and depending on your criteria.

But I do think if you just want an all-in-one model, you don't want to think too much about what you should use, then I think Opus 5.5 as I'm sitting here right now is probably the pick for a lot of people. And if you're on GPT, would you switch to Claude just for this? I don't think so. I think Astra is still pretty good. I don't think it's worth necessarily making a huge switch just for this, but if you're an AI nerd like me, then then it's probably worth testing.

And you probably already have a Claude subscription. All right, I am starting to see some results from the Fable test. We'll get into that at the end of the video. It's just wrapping up now, literally like 30 minutes later than the Opus version, which I think says a lot. Um, a Jazz here says, "Ananthropic have very likely sold for RSI. Have you noticed the newest models are not only smarter but smaller and cost less?

This is probably the result of distilling a much much larger internal model, aka Mythos. That's the way the Frontier Labs can compete with open-source cheaper trade-off. They'll just withhold the best models and pace via distillation." And it honestly makes sense. And Ryan said, "After seeing Opus 5.5, I refuse to believe Anthropic doesn't have something completely insane cooking behind the scenes. OpenAI apparently considers Bell AGI internally and it has already resolved 100 plus long-standing problems including Navia strokes.

Meanwhile, Anthropic has been remarkably quiet about what they have behind closed doors. What the f are they cooking in there? And you know, it's hard to explain, but so far using Opus, it does feel like it has that little bit of magic to it that is hard to quantify. If you are going to make the switch to Opus 5.5, this is actually quite a smart prompt. So, give it your project files and session history. So ask an agent to find all available local sessions for a project from Claude Code, Codeex or other tools that you use.

Have it read those sessions alongside the current files, projects, instructions, and plans and ask it to review how you work. So the sessions will contain your corrections, failed attempts, decisions that might never reach project instructions. So if you get an agent, Opus 5.5 to scrape these files and scrape the old session history, the agent can look for problems. So instructions you keep giving, tasks you keep doing manually.

This is a great approach. It's essentially auditing your own work using the new models. And whenever a model comes out, as be as best practice, I recommend that you do this. Ask it to reference the session and files behind each finding. Have it check those findings against the current project. So this could mean updating a skill, fixing a script, completing an unfinished feature, or adding a scheduled workflow. Review the plan before it makes changes and then make changes to your session. or at the very least if you're if you don't have an active session that you're still working on, get it to create a skill to use in the future which fixes the issue that you were experiencing in past sessions.

So then you can start implementing these changes every single time you start a new project. I think it's a very smart practice, a best practice to do whenever a new model comes out because it can find and spot holes in your old projects that will make future projects more efficient. There's nothing more annoying with Claude or any AI for that matter than having to like constantly reexplain context and constantly say the same things over and over and over and over and over again.

Actually, in my school community below, it's a free community you can join if you want to level up in AI. I'm going to leave a prompt. So, I'm going to take this full prompt. I'm going to improve it based on what I did. And then I'm also going to take this prompt here from Lance, which is a similar prompt audit. And you guys can just get it for free. So, you can just put it into your Claude um to save you time. So, when I upload this video, I'll pin a post in the school community and I'll put this prompt in the pin post.

So, you can just join the free school community, put it into your claude, and uh it'll just run the audit for you. All right, I see some action on the left hand side of my screen here. Let's look at Oh, this is crazy. These numbers are nuts. So, I built a simulated engine with 150 trader bots with a real order book for trading. And I just wanted to see how quickly Opus would do it, how quickly Fable would do it, and how much it would cost.

Opus took 12 minutes, Fable took 42 minutes. So imagine how this compounds. This is for what, 150 traders. Imagine how much this compounds if you were back testing like thousands of trading strategies, you know, it it would compound massively. Cost on Opus, $3.61. Cost on Fable, $17. Turns on Opus 24. Fable 62 and the outputs on both are essentially identical. It's a simulated order book. So the main reason I was doing this test is to see with a trading application, which is something I use AI a lot for, how much better would Opus be?

And I think already I am noticing Opus is much cheaper and much faster. Right now I'm running a back test as well, but it's literally going to take 5 hours. So it'll be on my finance channel once it's finished. I would also presume that Opus is going to finish a lot quicker and that's going through thousands of strategies um versus Fable and it's probably a lot cheaper as well. So I think so far it's clear just a smart cheaper faster bit of magic in this one.

I'll use the audit prompt in the school community if you want to use that in your Claude and I will see you in the next video where I'll continue putting Opus 5.5 to the test. See you in the next one. Peace out.