Accelerate your development with the Gemini API

Key Takeaways

[Applause] Hey folks, good morning. Uh, it's a pleasure to be here with you all. I'm Luca Martins. I'm a Brazilian. Uh, I'm an AI developer advocate at Google Duke Mind and I'm here with my friends first. Thank you, Luciano. Hi everyone. I am Tishta Basu Malik. I'm the product lead for the Gemini developer API. And looks like I should have brought an Indian contingent here. Thank you. Yay. Okay. So, the idea of this conversation is uh we want to share with you some of the new things you have available uh to develop your solutions using Gemini models and the Gemini API. How many developers we have here? Amazing. Okay. So um we can start talking about the Gemini models universe. We started Gemini by the end of 2023 and since then we have we have done a lot of work together between many different Google deep mind teams and by now we we are very proud of the points we we got in and all the stuff we launched during IO. Just doing a quick recap, one of the key differentials of the Gemini is that it is multimodel from scratch. Like since when we started developing the model, it always handled and understood multimodel data. It means that you can work with Gemini on with any kind of any format of information you want from text, image, video, audio, code, anything. Right? That's correct. and just a quick overview of what are all the types of models and what are all the families of models we have available. So what we're going to do today is uh Luciano and I have split this talk into two parts. So the first part of the talk we'll be talking about the models and then the second part of the talk we'll be talking about the API what are the capabilities and functionalities available through the API and as part of that we have a section where we do do a deep dive on agentic capabilities in the API. Yeah thanks for the ad. Yeah so starting with what are the families of models we have available. So the most uh you know perform uh performant and powerful model that we have is the Gemini 2.5 Pro. It's in preview right now. A few weeks ago we released a updated version of 2.5 Pro and it's suitable for highly complex tasks that require a lot of deep reasoning. I mean and coding has been of course a standout use case of that. Um we also have 2.5 flash which this Google IO we released uh uh you know a pre new preview version just yesterday for 2.5 flash and for 2.5 flash is our model size that is that has probably one of the best price performance ratios in the market today. That's right. Um we also have 2.0 flashlight. So this is a small fast cheap model that you can use for high volume tasks like summarization and then Luciano do you want to talk about the nano and the embedding models? Yeah absolutely uh we heard a lot of feedbacks from you folks and many of you are developing uh mobile applications or applications that that must run on devices. So then you have also available what we call the Gemini Nano which is a a smaller version of the Gemini which is able to run locally on devices like on Android devices if you are developing with uh the Android Studio using the AI core and also many of you are trying to create um applications that do some kind of semantic ranking or to organize information in large scale. So you have also one model that we call we call the Gemini embedding which the key objective of this model is to let you ingest text and then the model delivers you high quality uh multi-dimensional embeddings but how those models are behaving when we think about the key uh standard benchmarks. Yeah, we'll go to the benchmarks in a second, but the message that I want everyone to take away from this slide is whether it's a super complex task or whether it's ondevice processing, you now have a Gemini model that you can use for it and it's pretty powerful and it's pretty price effective. Yeah, maybe it is also worth mentioning Shesta that some folks here may be developing with Gemini since it was released. Yeah. So using the AI studio, using the Gemini API, the the the older versions of Gemini like 2.0, 1.5 are still available, but we really encourage everybody to start experimenting with the newer newer ones with better capabilities and better performance. Right. That's correct. So 2.0 Flash continues to be one of our most widely used models just again because of how good it is for the price it's at and we have a lot of people who are still using it. Um, and then we also have some people on the 1.5 models, but you know, we're encouraging people now we're two generations ahead. So, start using the 2.5 series of models and give us feedback. Nice. So, please tell us a little bit about the benchmarks first. Yeah. Next slide. Yeah. So we're not going to spend a lot of time on the benchmarks, but the two key points that I'd like to hit here is number one is one way in which we measure how good these models are is by putting them on Ella Marina, which is where developers and builders like yourself test out all these models and give it an ELO rating. And we're of course very proud that right now you have three Gemini models in the top 10 in the LM arena. including the top one including how did I forget that we are number one guys and we're very proud of it. Uh thank you. Um one of the things one of the use cases that has been really emerging in the last couple of months is of course app creation and vibe coding and going from zero to one. The best uh leaderboard right now for that is the webdev arena and we again are number one with 2.5 pro on web de dev arena. So if you're trying to build zero to one or few short apps, Gemini 2.5 pro is a great model for that. But in addition to user preferences, we also like to measure how well our models are doing based on rigorous academic benchmarks. And so I'm not going to go into a lot of details on these slides. These are available in blog posts. Reach out to me and Luciano if you want to know how exactly these models are doing. But the TLDDR here is whether it's in highly complex domain specific questions, whether it's in coding, whether it's in multimodal understanding. The Gemini 2.5 Pro is leading on a lot of these academic benchmarks as well. Finally, performance is one side of the coin. The other side of the coin is of course price. This is a chart from Swix that's been, you know, very popular on Twitter for the last few years on X, I'm sorry, for the last few uh months. Uh, but this really shows that the even for price performance, the Gemini 2.5 models are really at the frontier. Luciano. Yeah. And uh I I think it is worth mentioning uh Shesta especially for those of you who watched the uh Sundar's keynote yesterday and also the developers keynotes. We uh inside the deep mind we are doing this huge effort of trying to cover the different aspects of experiences you may have. So we we we will talk a lot about Gemini here but also we have the the model family that we are calling J media where you are able to to generate high quality images high quality videos. You also have and there is a colleague from our team Paul Ruiz who is delivering a talk this afternoon related to the Gemini robotics the Gemini family of models related to people applying uh robotics uh developing um robotics solution with applied AI and also just after this session we have one another one with Omar and that's right Gus Martins to talk about Gemma 3 which is our open uh the the open models family for for from Google deep mind. I also develop it with all the technologies we use to develop Gemini and some really good benchmark performance as well as Omar and Gus will tell you. Okay. So, uh still talking about the different the different experiences you you may have. We heard a lot of feedbacks from from you folks around the world and one of most asked features we always heard were related to having uh uh letting Gemini to to create high quality audio. So yesterday we we launched what we are calling the Gemini TTS model where you can from text generate high quality audio. But the coolest thing about the TTS is is that you are not just creating audio. You have abilities to customize emotions on the audio. You may have multiple voices. You can uh you can create the audios with different languages like Brazilian Portuguese and you can even create multe uh interactions. So if for example you want to deliver an experience where your users will be thoughts about something on a podcast like formats like the one you have on the notebook LM that people loves who who here is using notebook for anything amazing. Oh, so you can have like something like the AI overviews uh developed and delivered by the Gemini TTS model using the TTS, right Tesa? That's correct. And I think it is not the only thing related to audio. We brought that to the live interactions as well. Right. That's right. So the TTS models like Luciano said are available via chat endpoint. Uh and then on our live API yesterday, we also rolled out the native audio output models. So these are this is now Google's first release of an audiotoudio architecture. Um and there are two variants of this model that you have access through the Gemini live API which is our real time API. So you have the native audio dialogue model um which is uh you know the real benefit of the native audio output models are that the voices that come out of it are much more natural and much more compelling. Um the the native audio audioto architecture has better contextual understanding of what humans say to the AI. Um it can seamlessly transition between different languages. So maybe from Brazilian Portuguese to Bengali if you want to switch in that in the same language in the same sentence. Um and it's available on our Gemini real time API. We also have a version of this model available again on the real time API with thinking enabled. So for more complex use cases. So, one use case that comes to mind is let's say you're building a gaming agent to play a strategy game and you want that agent to be able to think but also to talk to you with relatively low latency. Um, we have a thinking version of the native audio model that's also now available from the live API. So, please try it out and give us feedback. Perfect. Um, so how people can start using those models first through the API of course. But before we go there, there are two other things that we should mention. It's hot of the news. You've probably all heard about it. We've also enabled an advanced reasoning mode called deep think. And the idea behind deep think is that you can uh uh you know the the model 2.5 Pro can think through various possible answers to a problem before giving you the best answer. Uh so that's going is available in trusted testers but will be rolled out more widely soon. And then we also have Gemini diffusion which is uh you know the first time we are uh testing a diffusion architecture as opposed to an auto reggressive architecture and these models are faster than the fastest model out today and for almost similar the performance. So try out both when you have access to both Gemini deep diffusion and deep think. Perfect. All right. So the Gemini API. So yeah uh if if you're site thinking especially after the keynotes yesterday there are many surfaces where you can interact with the Gemini models. So you have the Gemini app you have Gemini available on Google cloud as the code assist the the assistant on the the cloud assist inside the console. Uh you have a Gemini on workspace. So when you are talking about developer experiences, we have the Gemini one specific Gemini API the with similar experiences as you have on the no code UI that we call AI studio where you can start experimenting and developer uh developing your solutions using the Gemini models. Right. That's right. So basically the Gemini API is a very low barrier entry for having Gemini uh programmatic experiences. You have access to all public models there from the J media ones, Gemma, all the Gem the Gemini models variants. Uh you have a very generous free of charge uh cheer where you can start experimenting like just after this session without concerns about credit cards, costs, billing, anything of uh like that. We we are developing uh we keep developing more SDKs for this for for this API. So for now you have available uh one SDK for Python, JavaScript, Go uh and Java. We launch a Java during IO right asa and also you have the ability to use the API on other developer um developer tools you may be using like for example if you use Firebase Studio the API is available there. If you are a Google collab user, you have ways to interact with the G Gemini models as well. So the key idea is to make it easier for you. No matter where you are having your development experience, we are trying to bring Gemini API closer to you. That's correct. And I think this is also a good place to make a plug for Google AI studio which of course everyone knows about and loves. It is the place where a lot of our developers first test out the capabilities of the API before committing to building applications at scale with the API. And Google AI Studio now has code generation also available. So this is a um a quick high-level overview of all the components that are available through the Gemini API. It's pretty standard. You send in your prompt, you get a response back from the model. If there's a tool call needed, we do support a set of firstparty Google tools as well as function calling. Um, so in terms of the firstparty Google tools that we support, uh we of course uh support Google search as a tool. Um the best search engine out there is now available as a tool for the model to call when it needs to. Now, in addition to the information that you get out of search, if you want to retrieve more indepth content from web pages for applications such as you're building your own version of a research agent, so you pull some information from search, you pull a set of URLs, but if you want to extract a little more in-depth content, we just released a new tool at IO called URL context. And then of course you can take all that information and if you want to create beautiful charts out of it or run some analysis we make code execution available as a tool to you via the API as well. Then of course you have function calling and as part of all the improvements that we've released um as at Google IO we've also put a lot of effort into making sure our structured outputs uh functionality which lets you get outputs and JSON schema is much more robust and comprehensive. So again try it out and give us feedback. And then finally uh we do have a set of safety and copyright filters. Um these are configurable um by developers. Um the idea is in order you have the tools you need to make the applications you build safer and you have control about how for most of the filters you have control about how how where you want to set that threshold. Perfect. So as a PLDR if we could explain on a tweet how that work with the Gemini API and using the SDK you can work with all information you need doesn't matter on which format it is if you have like spreadsheets docs PDF files videos audio uh live interactions via voice and video right stuff if you need to give specific instructions to the model you have the ability to bring system instructions to the model follow on every interaction you have during uh your application usage and you have as results as or as outputs the ability to have generated text or generated images or audio or video or even keep chaining with more API calls using function calling. That's correct. That was a huge tweet. That was a huge tweet. Maybe threaded tweet. All right, so let's talk about some key Gemini API features. Like we don't have enough time in this talk to go through everything, but Luciano and I thought we would hit some of the highlights that we really want people to know more about and use. So one of the areas of feedback that we used to get a lot is uh you can of course upload files to the Gemini API. You can also if you're less than 20 megabytes, you can also pass some of these media information in line. But now you can also, and we've had this feature out for a few weeks, you can also pass a YouTube link and the Gemini API can analyze that information. Um, when you pass media files like videos, depending on how much you want to fit into the context window, you have now a choice between three resolution settings. At the lowest setting, you can uh you can process up to 6 hours of video, but then you also have more high resolution settings that you can use. Uh, we support dynamic frame rate per second in the G Gemini API. Um you can read that more about that in our documentation or come talk to me and Luciano. Um we support video clipping that's a new feature. Um and we support image segmentation. One other point I want to make on multimodal understanding is even in the days of 1.5 and 2.0 the Gemini models were some of the best models out there for multimodal understanding. Um, you know, the the example I like to give here is I was in Costa Rica and my guide showed me this is night and my guide showed me, oh, there's a glass frog somewhere there on a branch. I took a photo, but I did not see the glass frog in real life or in the photo, but I passed it to Gemini and Gemini not only saw the frog, it identified the species correctly. So, that's how good multimodal understanding is on the Gemini. Uh, please try it out. Uh we also support long context. Um um we and then we be with along with long long context is we have some of the largest context windows out there. Uh so like the equivalent of depending on whether you're using a model with uh 1 million context or 2 million context like you can read the equivalent of like eight novels um have the model read it and or entire code bases. Um but what can sometimes happen with long context is that your input token pricing goes up but for that we now have context caching. We have been supporting what we call explicit context caching for a while. Explicit context caching is when you tell the API cache this context and reuse it for the next uh few turns and you were going to get a 75% discount on the pricing. But now what we support is implicit context caching. So if we feel that you have a context that you're that is getting repeatedly used, we will automatically cache that for you and pass the price savings to you. So again, hopefully this is a huge price benefit to our developers. Um and then we we of course provide you with transparency to see how many tokens you're using. That's amazing. Yeah. you want to talk a little bit about text generation? Yeah. Uh I think one of the uh one of the greatest things that we we we are building step by step with the Gemini models and the Gemini API is you may have different experiences as Festa explained with videos and clipping specific time offsets or changing the frames per second. And then the first the the first like out of the box output of Gemini when we launched it in December 23 was text generation. So we keep having this as one of the choices you have. You can you can have an experience where Gemini will just create text and by text it may be like anything. It may be one structure JSON output or it may be some coding in Python C or any language of your preference. and but then you uh we we keep increasing the the semantic understanding capabilities of the models. So we are not just uploading one PDF file with a lot of text and some charts and some conclusions. You are also counting on the Gemini ability to understand how the information on those documents or on those spreadsheets connect to each other. Especially if you have a more complex situation where you are sending multiple PDF files or PDF files with spreadsheets and videos and everything without a huge head heavy lift from your side. You can count on Gemini to understand what's the message, what's the information, what's the reasoning behind all that information and do the math or do the understanding for you. Right SA? That's correct. Uh I just want to give a special call out to two features here which again developers have been asking us for one is you can get bounding boxes through the API and secondly in addition to bounding boxes we you have what's called image segmentation which is you get the bounding box information for an object in the image you get uh a classification of what that image might be. So that's sent to you as metadata and then you get this mask segmentation um of that specific area and that object as well. Nice. Um we also support streaming and you can set system instructions on top of whatever system instructions we already have in place. Yeah. And I think that's the same for the J media models, right? So, so now uh you have basically three main doors to use in media models. You have the imaginary the Google deep mind model with the the best high quality image generated. Yeah. You have one one variant of Gemini that we we call Gemini image out that lets you also generate images but with two key differences from imagine 3. First, you can have interled outputs including text and images together like having one explanation on step-by-step guide to do some action including visuals and also you can edit images right so you can have one first version of the image generated you want to change the shirt color or the background or add glasses to the character on the image you can keep chatting with the model asking to modify the image and you keep enhancing and optimizing your result vice versa. That's correct. And through VO, uh you all saw the release of VO3 that'll probably come soon to the API. But with V2, we support text to video and image to video. Um this is again from a recent blog post that our researchers put out about a week ago. Um you can see that 2.5 Pro also as I mentioned earlier leads on key video understanding benchmarks like MMU. So that's uh you know that's also something that we very proud of that we've been pushing the boundaries on. Okay. So now let's talk about the live API. So a lot of our features as I mentioned are available through the chat API but we also provide our the Gemini live API which is our real time low latency API for use cases that require more of these interactive realtime kind of experiences. Uh there are two architectures now available through this live API. There is the cascaded architecture which has native audio input but the output is done using uh texttospech um and we were using the same text to ste speech models that was used by notebook. A lot of people prefer this architecture because it's more reliable. We've had it out since December. Um and we're aiming to bring the latest 2.5 flash model to this architecture. We also starting IO now have the audiotoudio architecture that I mentioned where you have native audio input as well as native audio output. Um and this of course native audio output as I already mentioned gives you much more natural sounding voices and the ability you don't now any have to specify a language you can seamlessly switch between languages and flow in and out. We support tool chaining in the live API. We have supported this since December. So all of the tools that I mentioned, search, code execution, URL context, function calling, you can layer these tools in the same prompt um and get much more compelling results, get data from search, do some analysis, get the output. Um we have voice activity detection. Of course, we need it in the live API. What we now provide for you though is the ability to configure the thresholds. Um how much of time do you want after the end of speech to decide that the user speech has ended? Like that's one of like four thresholds that you can now set with the live API. Uh you can also disable our voice activity detection model and bring your own session management. We have a lot of parameters out there. So in its most basic state, the live API currently supports about 20 minutes of audio and about 10 uh a few minutes of video. But we now have various techniques for you to increase your session length, including sliding window, the ability to change resolution on what video is passed, the ability to decide um you know, do you want vid audio to be streamed only when do you want video to be streamed only when audio is being spoken or even when audio is not being spoken? um and other parameters that you can use through the live API. Um ephemeral tokens are coming soon. Um but that's one way to do authorization into the live API. And finally with the native audio output uh specifically the audio to audio architecture, we are also releasing a couple of modes for you to try out. One of them is proactive audio. What this feature does today is it lets the AI decide when to respond to you and when to when the whatever the human is saying is irrelevant. So imagine if I'm having a conversation with the AI and Luciano comes to me and says something unrelated, the AI will know not to respond to that audio output. So the AI proactively decides not to respond. So that's we are calling it proactive audio because we aim to bring much more proactivity to this feature. Um and then effective dialogue lets you pick up on the user tone and sentiment and lets the AI respond appropriately. Uh as I mentioned you also have thinking available with the live API. All right time for agents. Excellent. So, how many of you are trying to experiment solving your your computational problems using agents or multi- aent solutions? Yay, everybody. Brazilian. Okay. So, with that in mind, we always try to develop the new tools and the new capabilities of the model, thinking how you folks can use them in your uh in your projects to make agents better and more trustful. Right. That's correct. So if you start thinking about how our regular agents architecture work um what what do what do you have there? So basically we have three main blocks. Yes. We normally see one or orchestration layer, one models layer and one tools layer, right? Yeah. May I build upon that? Um I think as we've been mentioning with the 2.5 series models um these models are predominantly trained to be really good at planning and reasoning which when you think about it is a key part of makes it what makes an agent work. Um so the model layer is where uh a lot of the prowling and reasoning happens and then of course there's the tools layer. uh we've already talked a lot about we have a set of firstparty hosted tools from Google Google search code execution u and URL context uh as well as a few other tools some of you may have heard Sundar mention the computer use uh tool that's coming so we're going to make it available through the API we've already rolled it out into trusted testers but that's going to be publicly available soon and then we have a couple of other tools on the way as well right and then there's the orchestration layer as you were saying Luciano. Yeah. Yeah. Absolutely. So basically when you are creating an agent you want to to give like key directions to these agents. So basically how it's going to behave what's the profile of these agents uh what what's the goal of it. It's going to help with researching with coding with learning any specific area you are trying to to address um to address actions for for your users. you you need to count with some memory for these agents. So you can keep like the previous users interactions with the model or you can try to extend this agent memory using any mechanism like rag or adding PDFs or using the long context and also you you must count on the model to do the key reasoning part of this agent. So how to bring all those stuff together, how to give the best answer to the users and when and how to use each of the tools that are available to these agents. That's correct. So yeah, so basically uh if you put all that together, we are looking for applications that you are calling like multi aentic applications or multi- aent applications where we want some autonomous integration with our tools. We want those those agents to take actions to help our users. We count on the models like the Gemini 2.5 Pro, the best one on reasoning capabilities to reason and plan all the actions to be uh to be given to this user from like helping with chop shopping to generate new codes for a new application. This model must be able to be continuously learning. So if you want to append more pieces of information, fresher information, use like uh tools like the Google search grounding righta to bring up to date information from Google search. This model must be able to support that. And at last but not least, we count on multi- aent collaboration. So we must be able to not create huge monolithic agents but also to keep connecting to other specialized agents to bring a better experience. Right? That's right. Uh you've probably heard us talk about all of these components 50 times already in this talk. So I'm going to go through the slide really quickly. uh but one thing I we wanted to emphasize is when we started thinking about how to enable agentic capabilities through our API we made a conscious decision on first focusing on providing high quality primitives and once we had made some progress along that that's when we have now started to do things like release MCP through our SDK uh and you'll see some more higher abstractions rolling out in the next couple of months Uh but in terms of agentic primitives again uh we have the 2.5 series models which are thinking first models. You now have deep think which is an advanced thinking mode on top of 2.5 pro. Um and with flash today and pro soon you have the ability to set budgets. So you can tell the model when to think and how much to think. And that lets you both control cost, latency, and whatever is the amount of thinking that's suitable for your application. In terms of API, depending on what agent you're building, you may want to use different versions of the API. So, if you're building a research agent, maybe you want to use the chat API. If you're building a gaming agent, maybe you want to use the real time a API or a customer support agent. Of course, there's um there's um you know, you can build research agents also on the real-time API, but in general, you now have these two APIs. Um and then we've talked about tools a lot. So, all I'll say is give us feedback in terms of what are some of the other tools that you would see Google make available through the Gemini API. Yes. So as you just said the Gymnast 2.5 Pro modu uh 2.5 models are thinking capable so they can do more advanced and more uh complex reasoning. Yeah. And maybe we could highlight that now specifically for the 2.5 flash we have the ability of using what we just mentioned the thinking budgets where you can calibrate how deep and how much the model will do reasoning cycles and also for both models you can you can have access to what we are calling thought summaries. Yes. So basically you are not only counting on the model ability to reason but also you can see on your response which key steps or which key reasoning uh thoughts the model had to get to our conclusion. So basically that's a pretty basic Gemini API uh code uh using the Python SDK. Uh for those of you who never used the SDK before it's pretty straightforward. So three key blocks you need to import the SDK you instantiate the client and then you'll interact with the model you want. Here uh we are interacting with the Gemini 2.5 flash. You send your prompts and then you have one specific bit called thinking config where on this code you are asking the the API to include the thought summaries on your response and that's the include thoughts equals true and also the amount of tokens you want to use uh during the reasoning process. on your response uh you're going to have parts of the response will be the thought summaries. So uh it's separated from the final answer by intention if you want to keep those on your on your back end on your data warehouse your BI environment you are free to do that and also you have a second part which is the final model answer that may be like for example the part that goes to the to the end user vice versa that's correct so here we added like two quick gifts so showing how the the thought summaries work you have on the on the on the left gifts the AI studio experience and on the right the Google collab experience and pretty much we ask it the same question and you can see how you can interact with the uh with the thoughts on both environments right okay so what about the enhanced tooling we have now I think in the interest of time we can skip the slide because we've talked about tools a lot the one point I do want to make is I mentioned when I was talking about the live API I that one of the things we started doing in December itself is we allowed you to chain tools. So use multiple tools together in the live API. We're now rolling that out to the chat API as well starting from search and code execution together. But you'll see more the ability to do more combinations of tools in the chat API as well. Amazing. So basically uh that's what what you just said. That's a similar experience with the SDK. Now you have two tools included on your interaction. You have the in this case the code execution tool and the Google search tool. And then on your answer you can see also on separated structures the code execution results including the Python code used by the code execution and also the output of this execution. The Google search grounding results including the Google search real search that was performed and the results of this and your final your final answer from the model. Right. URL context. As we said, this is a new tool that we are rolling out. Uh it's one of the tools that powers deep research if you've used Google's research agent through the Gemini app. Uh and the idea is um given a set of URLs we allow you to extract more in-depth content of course in a way that's approved by and respectful to our publisher ecosystem but you can get more content out and this is really helpful for research and analysis type of use cases. You can use this tool by itself or you can use this in tandem with the search tool. Yeah. And again a quick uh code snippet. You add the tool called URL context and you can send directly on the prompt without further reports all the links up to 20 links you want to consider on the model processing on the model execution and you have all the reasoning all the semantic extraction performed by the model really fast. like some people were still taking pictures. But if you all missed taking an image of a slide, all of this information is available in our documentation. Absolutely. And Luciano and I are also here to answer questions. Uh function calling I think um we've had of course function calling is bread and butter for all kinds of agent agentic applications. Uh the Gemini API has supported single function calling, parallel function calling and compositional function calling for a while. Compositioning. Compositional meaning where you can put a whole logic around if A happens then call function one. If B happens then call function two. As part of IO, we are also releasing asynchronous function calling through the live API. So imagine some tasks like you start having a conversation with the AI agent but in the background you ask it to crank out on some other task and that task may not be a realtime task but it may be something that's relevant to the conversation like maybe analyzing a lot of context about the user to provide some answer. Um you can now turn on an asynchronous function. The the uh the thing uh the function will do its thing in the background and then when ready it will notify you with the results. Amazing. And maybe one of the greatest news for the Gemini API SDK during IO is that we brought we heard you many of you gave this feedback for us and we added to the Gemini API SDK support for MCP. So now you don't need to have like separated and uh and uh siloed codes with MCP clients, MCP servers and then your Gemini API interactions. You can have it all together using the same code base of the Gemini API SDK. Very exciting. Yeah. And and and thank you. And still talking about like those interactions, we know that we have many choices of agents frameworks. You may have heard about the Google ADK, the agent development kit launched during the Google Cloud Next few weeks ago. We may have heard about the agent to agent protocol as well. You have others uh other choices like link chain, lingraph, everything. and we keep open collaborating with all those internal and external open source efforts to bring the best the best experience for you all. That's correct and especially uh thanks to folks like Philip Schmidt and Patrick here on our dev developer relations team. uh we have been working on uh building closer relationships uh closer interactions, better code samples with some of the leading agent frameworks uh like lang chain crew um and so um that is also some of the areas where you'll see the Gemini API show up. Amazing. So if you if you folks uh are concerned about a lot of the amounts of stuff you we've just shared today that's a huge dump of things. Maybe the best the best suggestion or the best guidance we could give you is first asa mentioned we have the Google deep mind uh deing distributed across the globe. We have people in Latin America in Europe in Asia uh we have a huge presence here in the US. So please count on us, connect with us online on on onsite events like this and think on on like those top six uh actions when you are creating your aentic experience from having clear objectives on your mind. Try to laser focus on on the problem you you need to solve. Do a lot of interactions on on your development. do live coding if you need to uh per uh peron do vibe coding if you need if you are working with something you are not that familiar count with the Gemini 2.5 models to help you with your coding experience and always always focus on your user experience uh that's that's the the best key for success you may have with your tools right the OG rule of product absolutely so yeah and that's it I hope you enjoyed that start building now. We brought you here. [Applause] Thank you. We we we give you some links here. So, some of you may may get the reference, but basically the first link is the AI studio, the UI and no code experience where you can experiment all the Gemini models features really fast without writing a single line of code. The second link is the Gemini API docs where you may find all the things we shared here and a way more of the other features and possibilities with the models. And the last one is the Gemini cookbook, a GitHub repostory that our team curates and keep updating with a lot of sample experience for you. Thank you so much. Thank you all for coming out to hear us. [Applause]