---
type: transcript
date: 2025-05-06
youtube_id: mZNLegBg8BA
---
# Transcript: Gemini 2.5 Pro 05-06 in 6 Minutes: The Best Coding Model?
Google has just released an updated version of Gemini 2.5 Pro which makes it even better at coding. In this video, I'm going to go over the announcement. I'll look at some of the benchmarks and then I'll also show you on how you can get started with leveraging the model. Right off the bat, in terms of some of the key aspects of this model, you can expect meaningful improvements for front end as well as UI development alongside improvements in fundamental coding tasks such as transforming as well as editing code and creating sophisticated agentic workflows. One notable piece with this announcement is their leap from the previous version of Gemini 2.5 Pro. This model was originally released on the 25th of March, just over a month ago. And with this latest release, this is the first model to outperform Claude 3.7 sonnet on the webdev arena. The leap from their previous model, which just came out, is an ELO increase of 147. So from 1273 all the way up to 1420 with Sonnet 3.7 sitting at 1357. In terms of being able to access this model, you can head to aistudio.google.com to access this completely for free. You can also get your API key here. Now, the one thing to note with the model, so it does have a million tokens of context. This model is ranked number one on the LM arena. In terms of the preferred responses across all of the latest Frontier models, we can see this ranks even above 03, which just came out last month. Another thing to note with the model is while coding is the focus of this release in terms of a lot of the examples that you'll see out there, we can see across the board that this model does outperform from everything from coding to math, creative writing, instruction following, so on and so forth. Another great thing with the model is the pricing is exactly the same as the previous generation. If I look at artificial analysis, an independent analysis platform of different AI models, just to give you an idea, in terms of the blended rate for the pricing, it is $3.40 40 cents per million tokens of input. When we compare that to Claude 3.7, we can see that's almost half the price of that model. Now, more specifically, the pricing of the model does depend on how much context you feed it. While you can give up to a million tokens of contacts, if you're only giving it up to the 200,000 token mark, that's going to cost $1.25 per million tokens of input, $10 per million tokens of output, and then for all the tokens that exceed that 200,000 tokens, it's going to be $15 per million tokens of output. Now, additionally with this model, you do also have the ability to see all of the thinking traces. Now, one thing to note with this model is it is very fast, especially for a model of this capability. It does seem like it is likely in line with the same speed as the previous generation of Gemini 2.5 Pro. Given that this is a Frontier model, the speed of the inference does appear to be quite fast. So, in terms of some of the tools that are built into the Gemini Studio API, you can leverage structured outputs, code execution, function calling, or additionally grounding with Google search. Another way you can access the model is at gemini.google.com. You will be able to try this out for free. Now, if you do have Gemini Advanced, their $20 a month tier, you will be able to get considerably more requests. I was rate limited fairly quickly, I believe, within a few prompts of trying to generate a web app. So, within the blog post, they had a number of different examples. One that I actually found pretty interesting was this video to code example in the starter apps within Google AI studio. If you go to Google AI studio and you click on video tolearning app, what you can do is you can paste in a URL and what it will do is based on the context of what's within that video, both the visuals as well as audio, we can see these different dynamic applications that were generated as a result. There are a handful of different front-end web apps that you can go and check out within here if you're interested in trying this out. Now, in addition to the video examples, here are a handful of examples of passing in an image as well as a prompt. So, we can see some various generations of what the model does with these different examples here. In terms of some reactions, they did include one of the founding members from Cognition, the team behind Devon. And they mentioned that the updated Gemini 2.5 Pro achieves leading performance on our junior devals. It was the first ever model that solved one of our evals involving a larger refactor of a request routing backend. It felt like a more senior developer because it was able to make correct judgment calls and choose good abstractions. Additionally, the president of Replet mentioned that we found Gemini 2.5 Pro to be the best Frontier model when it comes to capability over latency ratio. I look forward to rolling it out on Replet agent whenever a latency sensitive task needs to be accomplished with a high degree of reliability. Here's an example that Demis, the CEO of DeepMind, shared where it's a simple doodle of a web application. And basically, with just the prompt of can you code this application, the model is able to generate this working web application that you see here equipped with all of these different color selectors as well as the ability to turn the drawing into a sound. We can see here that based on the instructions within the photo, we have the different color brushes at the top here. We have the doodle area and then finally we have the button to turn the drawing into a sound. Again, this just gives you another idea in terms of what the model is capable of. If you do leverage this within the Gemini web app, you will be able to leverage their canvas feature. And what that allows you to do is you can pass in both images, videos, as well as text be able to generate these web apps. And one of the great things with Canvas is you will be able to preview it directly within the browser. You won't need to take that code that it generates, paste it within a code editor or anything like that. You'll be able to see all of these generations and try out that new web development capability of the model without having to have too much friction. Finally, I saw a few other fun examples online. Here's an example of Gemini 2.5 Pro generating the now infamous question of who will win a 100 men versus a gorilla. And here is a simulation of a web application. All in all, I look forward to trying out the model in something like Windsurf or Cursor to see how well it performs, but it does look very promising. Given the price, speed, as well as performance of this model, this does look like a very promising option. Kudos to the team over at Google for this release. And if you found this video useful, please comment, share, and subscribe. Otherwise, until the next
