Claude Sonnet 4.6 in 7 Minutes

--- type: transcript date: 2026-02-19 youtube_id: EUzc_Wcm6kk --- # Transcript: Claude Sonnet 4.6 in 7 Minutes Anthropic has just released Claude Sonnet 4.6, the most capable Sonnet model to date. Now, the one thing that they really did highlight within the blog post that I thought was interesting was in and around computer use. This model performs exceptionally well at actually using computers in real world tasks. There is a benchmark called OS World where what it does is it measures how well the model can accomplish real tasks in real software. Think things like Chrome, Office, VS Code, and more. And the thing with this benchmark is effectively to be good at this is you have to be able to interact with a computer in the same way that a person would. You have to be able to click around the screen, use the keyboard. And one of the things that I want to highlight is just how far we've come in about a year and a half. We saw a new variation of Sonnet 3.5 came out with improved computer use capabilities. At the time, they also had a demo in and around computer use. And ever since then, this capability has really been on a tear. Now, the one thing that is really nice if you haven't tried it is they do also have a Chrome extension, which is a really good place to leverage exactly this. You can use Sonnet 4.6, you can ask it to do data entry within a spreadsheet. You can use different programs, different web apps. All of a sudden, your agents are going to be able to act just like people would. There's not going to be the same barriers where there necessarily needs to be an API to do something in terms of evaluation. So, this isn't the model that's going to leapfrog Opus 4.6 six or anything like that. In terms of a lot of the core capabilities, Opus 4.6 still is a better model for a lot of different use cases. But in a number of task, it does come awfully close to Opus 4.6, but there are some where it does outperform. So things like agentic financial analysis as well as office task. Then the one thing with the benchmarks is just consider there's no one metric that really tells the full story of a model. You have to take all of this in aggregate because if you have a model for instance that's good at agentic coding as well as good with office tasks all of a sudden you have a model that has a ton of general purpose capabilities. If you have a model that can use a computer for instance as well as write scripts build applications and do both of those things in tandem that all of a sudden becomes a very effective agentic model. Next up in terms of artificial analysis so this is a great company that does coverage in the space. They have a number of different benchmarks. They do also measure different providers and what have you. Now the one thing with this benchmark is as they state this is a primary metric for general agentic performance measuring the performance of models on knowledge work task from preparing presentations and data analysis through to video editing. Models use shell access and web browsing in an agentic loop. And this is their open source harness. What you can see is claude sonnet 4.6 six with adaptive thinking on does outperform all other models. And the thing with adaptive thinking that is quite interesting, this was just released with Opus 4.6. All of a sudden, the model can decide when to think harder on particular tasks in the moment. You're going to be able to dial up as well as dial down the thinking without actually being explicit in how much thinking to have throughout the whole chain of commands, which is a pretty interesting feature that they just released. Now, just to briefly touch on the model cards. There were a number of interesting things that came up within this. I'm just going to highlight a few. First, I want to point out is overly agentic behavior in guey computer use settings. So, it's a little bit of a mouthful, but effectively what they describe within this is a sonnet 4.6 is substantially more likely than previous models to take unsanctioned actions in the guey or when leveraging a computer. think things like fabricating emails, potentially even initializing non-existing repos, or doing things like bypassing authentication walls without asking. This was a similar thing that came out within Opus 4.6. But the good news with this is unlike Opus 4.6, this behavior is easily steerable with system prompts. If you do notice it go off the rails a little bit, just know that you can go back to your system prompt and give further instructions to have the model be a little bit more steered in terms of the direction of where you want it to go. Further, just to build on this, the one thing that they did highlight is in some simulated tests, Sonic 4.6 completed simple spreadsheet data management tasks that were clearly related to criminal enterprises in arenas like cyber offense, organ theft, and human trafficking. It would refuse these task in non-geuey scaffolds, surprisingly flimsy justifications, including a request to work with a set of password protected personal data files for a company despite being directly asked to do so and explicitly giving the password. What is interesting with this is when they were testing it, it didn't hesitate to do some very questionable things and it did refuse for something that should be relatively straightforward. If you're giving it the password as well as instructions to actually access your personal information, you would expect it to be able to do that. Next up, I quite like this benchmark. It's becoming increasingly popular is Vending Bench 2 from Anden Labs, which is a simulation where you give the model the capability to effectively run a business. It's going to be running a vending machine and all of the different ins and outs of actually operating that vending machine. It's up to the model to decide and see how well it performs. The one interesting thing and this has come up in some other tests with this is when it was tasked with do whatever involved in price fixing lying to competitors in the simulation. So it was comparable to Opus 4.6 six. But and the one thing that they call out is this was a notable shift from the previous model of Sonnet 4.5 where it wasn't quite as aggressive. So, I'm not going to go through all of this, but there are a ton of really interesting pieces within the model card. I'll link it within the description of the video if anyone is interested at taking a look. Okay. And last up in terms of the design sensibilities of the model just before the video I just gave Cloud Code a very simple sentence where I basically said build a full stack SAS application that leverages Nex.js and within here here is what it generated for me in terms of how it looks. So it's very similar to a lot of other models but there are some nice things within this. It's not leveraging favicons for instance. It's also not leveraging a ton of linear gradients which we'd also see within a lot of other models. There are some subtle elements of that, but in terms of the actual design style with just one prompt, it gives you a really good scaffold that you could potentially build off of. And it did also build this nice little graphic within here. In terms of the sensibilities, that's one thing with even models like 6 months ago. For it to generate something like this, it would look a lot more like slop is how people would refer to it. There'd be a ton of linear gradients everywhere. There'd be these weird out ofplace favicons. everything would sort of look generic and cheap where it's starting to move in a direction where things just sort of look a little bit more prim and proper. I don't think we're quite there with this model, but it is improving with each generation of these different models. But otherwise, I just wanted to do a really quick one going over Claude Sonnet 4.6. In terms of where you can access it, I'll put links within the description. You can access it anywhere. You can get it from the API. You can get it from cloud.ai. You can also get it within cloud code. Then the one thing to note is if you do want to try out the million token context window, you will be able to try this within cloud code, you can just do a quick Google search for the flag to try it out. But one thing to know with that is when I personally tried it right before recording this video is it quickly ran out of tokens. So while it's in beta, just be mindful that you might run into some limits with that. But otherwise, that's pretty much it for this video. If you found this video useful, please like, comment, share, and subscribe. Otherwise, until the next