Transcription
[Music] [Applause]
Hi, welcome to another video. While most people were waiting for Gemini 3, Google has launched Gemini 2.5 Computer Use. This is a model similar to what we've seen with things like Anthropics and OpenAI's models, but now it's by Google. You can pair this model with tools like Browserbase, Playright, and similar ones, and make it navigate the web to get things done for you or test UIs and stuff like that.
They have built it in collaboration with browser base and it performs the best on stage hand evils. Gemini was already really good on their evils and it has been really good for web navigation and similar tasks. However, now there's a model specifically fine-tuned for that. It is supposedly a fine-tuned version of the Gemini 2.5 Pro model, as this is also a pretty fast model. This is the model that has been used in things like Project Mariner, the Firebase testing agent, and AI mode as well. So that is really great.
In their system cards, they say that the model has primarily been optimized for understanding and interacting with web browsers. It has not yet been optimized for OS level navigation. So it's basically only for navigating and surfing the web. Many people were asking for OS world benchmarks for this, and that's why there are no benchmarks for it yet because they focused solely on web navigation. I think that's a very good decision because that's mainly what most people want to do with these models.
It also works similarly to Anthropic's API where they give you a specific tool in the API to use this feature. In the API, it isn't free and costs the same as Gemini 2.5 Pro, which is a bit of a bummer, and I'd have liked it to have a free tier, but BrowserBase has it for free, and we can check it out there. There's also their agent quick start, which is like a complete quick start tool to use this model. It's their reference implementation. To use it, you'll have to clone it, get into it, install the dependencies, set the Gemini API key, and then just run the tool by running the main file with the query that you want it to do, and it should go ahead and do it. You can also configure it to use a browserbased sandbox or similar setups as well.
In some time, it should also get support from things like Kilo, Rue, and Klein, where they can use these tools to check for UI components and test the apps they make. So once it's available there, it should be great. It's also a bit cheaper than Sonnet for small tasks, but over the 128,000 context window, it will cost the same as Sonnet, which can be a bummer. So, keep that in mind. It's still one of the best options.
Anyway, you can also test it on the browserbased site, and I tried some stuff with it to see how well it performs. The first question that I asked it was to solve today's Wordle, and this is a pretty hard task for most models. None of them are able to do it, and this was no exception. It also failed this task, which is a bit of a bummer for sure. It's not something that a lot of models can get through. So yeah, this is not that great, but it can still surf the web, do basic tasks, and that's what most of you would want to use it for anyway.
It will really come in handy for use with AI coding tools. It's just that there isn't any MCP or similar setup to integrate it with them yet. I really hope that Google would have come out with something or at least had it integrated into Gemini CLI, since that's their tool, but that isn't done yet. So, I'm waiting for it to get implemented into things like Rue and Klein, where it can work hand in hand with the AI coder to test UIs. Probably an MCP would be great. It's a bit tedious for coders to integrate this as well, because it's not like Sonnet where you just add the computer use tool and it works. Here it's a different API endpoint, and making different API endpoints work together can be a hassle. So there's that.
I think that Gemini 2.5 flash plus browser use is still the best option for automating web navigation and context gathering tasks. I only see value in this for checking UI testing and similar uses. So, there's that. That's mainly about it. I think it's a great model, but until it's implemented into something that's more usable, it's not that great yet. Something like AI mode is a good example of this model's proper implementation, but in the open community, there isn't a ton of good implementation for it. Once it's integrated into things like R's computer use, I think it would be pretty awesome that way.
When I was using Sonnet, which I'm not currently using, I used to have the testing tool and browser use tool enabled in Kilo because it was pretty good. It used to go to the site, check if everything was correctly implemented or not, and stuff like that. It worked really well, but it wasn't perfect either. OpenAI also did something similar with their computer use API, but it was never really implemented into anything I could use apart from their own tools. Even Codeex doesn't use it, which is surprising since it's their own project.
So, I've been thinking about these agents for a while, and I think most of them are still more of a marketing gimmick for now until they get really better and super fast because all of them get stuck, take super long, and just aren't that good. That's why I don't use Comet because it's mostly marketing. I can't find any good use cases for it either. So the only use case for me right now is testing apps and nothing else. It might be useful for data entry people and companies, but I'm still skeptical about it. And now I'm just waiting for Gemini 3.
Let me know what you guys think about it in the comments as well. Let me know if you guys know of any good way to use these models apart from testing. And if you guys use it for anything apart from the testing stuff, as I don't use it for anything, it can be cool for research and stuff, but I have seen that browser use with Gemini 2.5 Flash is generally better even without vision and way faster. So yeah, there's that. I really like these models, but unless they are super fast, I don't find any much use cases for these models. That is majorly about it. Overall, it's pretty cool.
Anyway, share your thoughts below and subscribe to the channel. You can also donate via superth thanks option or join the channel as well and get some perks. I'll see you in the next video. Bye. [Music]