Mistral Launches API for AI Agents
08:25 Share

Mistral AI, the famous French AI startup, has just launched an API. So essentially, developers can now use it to build AI agents. They're able to run Python. They're able to generate images. They're able to perform RAG, a ton of other features that I'm actually really impressed with. I'm going to be breaking down all of that on the podcast today and also showing you why I think you shouldn't count out Mistral. I feel like it kind of faded for a while, but it looks like with the latest feature update, they are coming back roaring and pretty much competitive with

the top AI models. And I'm really excited about that and the additional competition. Also, they have some interesting nuances and things they're doing that I think are going to make them more competitive than other AI models. So I'll be getting into all of that. Before I do, I wanted to mention, if you haven't already signed up to try my own startup, AIbox.ai, I would love for you to do that. We have Mistral AI on there. Essentially, AIbox allows you to access the top

top 20 to 40 AI companies all for $19 a month. So you don't have to have subscriptions to all of them. We have misceral AI on there. We have meta Google deep seek, anthropic open AI. And we also have a bunch of cool image generation models like a ideogram and flux. You may not have tried. Um,

and audio models. You can chat with all of these, all the different models in one single chat. So you can just switch mid chat to be like, all right, open AI. I don't really like your response to that. Anthropic takeover. I need a code problem solved. And I know, you know, Claude is just better at code. Boom. It spits it out in the same chat. Then you switch back to another one and continue your conversation. We've got a bunch of cool features. One in particular that I love is a media storage. So every single piece of media image or audio that you generate is over there. A lot of times I will generate like

a podcast thumbnail or something that I need. And a month later, I'm like, I really liked my prompt I used for this. I cannot remember what chat of my 50 chats down it was. And it's really hard to search through them all. So the media store just saves all of those. You just go scroll through, find the image that you generated. You can click on that image. And then when you click on it, it will tell you the exact prompt, the AI model. And there's a little button you can click on that will take you to the actual chat you are having

when you generated that piece of content. And yeah, you'll be able to see the context. You'll be able to see anything else that you might have done inside of that chat. So it's super useful. The link's in the description to AIbox.ai if you want to check it out. Let's get into what Mistral is doing.

So, Mistral is a company that I am excited for, not just because it's a French company and I secretly want them to succeed so that I can have an extra excuse to go to Paris and talk to them probably in the future. That is not the only reason I'm excited for this company. I actually just am excited for the competition and they have a bunch of cool features. So, Sophia Yang, who is over at Mistral, made an announcement about this. She said...

announcing Mistral AI agents API, a game changer and building AI agents and putting AI into everyone's hands. I kind of feel like whenever they make these posts, they have to use their model to make them. And so like, it feels like I'm getting a sneak peek into what the model is able to produce, because I don't know, it's just no one says a game chain game dash changer and building AI agents. And like, they probably just use Mistral to do without no hate. It's just funny on the on

on the thing. Okay. And also like all of the emojis in here, it feels very AI model-y, but maybe that's just me. Okay. This is what it's able to do. Code execution. The agent can run Python code in a secure environment. This is amazing. We've seen people like Claude and even OpenAI

having like code modules in there, but being able to actually run the code, I think is really impressive and very cool. So that is now in there and it's accessible through the API. So developers can use this and plug it into things. So I'm really excited about the code execution beyond just code generation.

Number two, image generation. This is something that they were, I think, lacking was kind of the image generation. A lot of other models, you had Grok going, just partnering with Flex for a quick and easy way to get an AI model embedded. You have Google with Gemini. You have OpenAI. Anthropic feels like it's slacking a little bit without their own image generator because now we even have Mistral with their own image generator in there. So it's like Claude and Anthropic are good for code, but...

It's like every AI model's got their strengths and weaknesses. And I'm excited that Mistral seems to be grabbing this image generation tool in there. They also have a document library, so you're able to access documents from Mistral Cloud, which is essentially using your RAG functionality with user-uploaded documents. So you can upload, let's say your company has a whole bunch of documents that outline how you do certain tasks.

You can upload all of those to a specific folder and then it can download, essentially can query against them. So you could be chatting with it and then all of a sudden it's, you know, you're like, hey, like my company's fourth quarter report, what was it exactly? And it will be able to go to the documents folder that your company has and kind of view those documents. So, you know, reg is a super useful feature they added. The other one that I'm excited about is web search.

So they essentially have the ability for you to access the internet and it's not free. I'll break down the costs of all of these in a second, but it's very useful. A lot of other people are doing it. Grok kind of ties into X to be able to grab information and Google obviously uses, um,

uses Google. And then we have Cloud, which recently added this as well. So I'm happy they're doing that. It feels like catch up. They have MPC tools, so flexible and extendable interfaces for agents to access real world context. No additional dependencies are required. I'll break down some of these use cases that they mentioned in a second.

But it has memory and context. This is something everybody else has, but essentially maintain conversation history and manage conversations flexibly. Opening Eye, I feel like it's gotta be six months ago or more whenever they released this. Maybe time flies. Don't quote me on any timelines. But yeah, memory is awesome. Grok and Opening Eye both do that. Agent orchestration. So you can coordinate multiple agents to solve complex problems really easily. So yeah, so a ton of really cool features. I wanted to break down a couple that they just shouted out themselves.

agents, quote unquote, that you can use that they have a kind of, uh, they have like a quote unquote cookbook where they show you exactly how they would, uh, run some of these agents. One of them is a coding assistant with GitHub. One is a linear tickets assistant. One is a financial analyst. One is a travel assistant. One is a nutrition assistant. These are kind of just samples of things that, that you can do with it to get started. The last thing I wanted to cover is, um,

The, just how much smarter the model got, it got significantly smarter with web searches. This is kind of a no brainer, but I think it was kind of the missing piece for Mistro because it felt like it was falling behind sometimes. So it's gotten significantly smarter. It went from like on the simple QA data set, its accuracy rose from 23% to 75%. This is a big jump. And the other thing I wanted to mention was the pricing for all of this. So if you're going to be using it, it's about 0.4 cents or,

Sorry, it's 40 cents per million input tokens and $2 per million output tokens. For the web search, it's $30 for 1,000 calls. Now you can imagine companies like Google probably can get this cheaper because they're working with their own infrastructure and Mistral, you know, it's a little bit more expensive. Code execution, it's $30 for 1,000 calls. Image generation, it's 100 bucks for 1,000 images, which honestly is not that bad.

Premium news access, $50 for 1,000 calls. Document library with RAG. It's included if you have their Teams or Enterprise plan, up to 30 gigs per user on some of the tiers. And then the last thing is custom connectors. So audit logs, SAML, SSO, and other enterprise features are available. You got to contact their sales team to get the pricing on those, but they're for the Teams and Enterprise. So

They're charging more money. They're coming up with some really cool features. Personally, I'm really excited. And I think this is definitely pushing Mistral quite far ahead as being much more competitive in the industry. I'm excited to see where they go with this. I'll update you as it seems like this is launched and we're starting to see Mistral added to more and more applications. I am very bullish on Mistral. I think it's a great company. It got in in time to be one of the top AI companies. So I think they're going to be able to...

They're going to beat out others that seem to be kind of falling off, like the inflections or others like that. So excited to see what they do. Thanks so much for tuning into the podcast. If you enjoyed the episode, make sure to leave a rating and review. Always appreciate them. And also make sure to check out AIbox.ai. The link is in the description if you want to try all of the top AI models on one platform for one price. Thanks so much, and I will catch you next time.