He’s models are withdrawing at a dizzying pace, by all of large technology companies like Google to Startups like Openai and Anthropic. Keeping tracks of the latest ones can be overwhelming.
Adding confusion is that models of it are often promoted based on industry standards. But these technical metrics often reveal little about how real people and companies actually use them.
To shorten the noise, Techcrunch has compiled a summary of the most advanced models of the one issued since 2024, with details on how to use them and what they are best for. We will keep this list updated with the latest omissions.
There are literally hundreds of thousands of models of him there: Huggingface, for example, host over 900,000. So this list may lose some models that perform better, in one way or another.
The models of the one released in 2025
Openai O3-Mini
This is the latest model of Openai’s reasoning and is optimized for Stem -related tasks such as coding, mathematics and science. It is not the most powerful Openai model but because it is smaller, the company says it is significantly lower. It is available for free but requires a reconciliation for heavy users.
Deep -open research
Openai’s deep research is created to do in -depth research on a topic with clear quotes. This service is only available with PRO subscription $ 200 per month. Openai recommends it for everything, from science to shopping research, but be careful that hallucinations remain a problem for him.
Cat
Mistral has launched versions of Le Chat app, a personal multimodal assistant. Mistral claims that chat responds faster than any other chatbot. He also has a paid version with up -to -date journalism by AFP. The tests from Le Monde found the performance of the impressive Le chat, though he made more mistakes than the chatgt.
Openai
The Openai operator is meant to be a personal practitioner who can do things independently, how they help you buy groceries. Requires a $ 200 -month reconciliation chatgpt pro. He’s agents keep a lot of promises, but they are still experimental: a Washington Post reviewer says the operator himself decided to order a dozen eggs for $ 31, paid for the reviewer of the reviewer.
Google Gemini 2.0 Pro Experimental
Google’s long -awaited flag model Gemini says it excels in coding and understanding general knowledge. It also has a super -long 2 million signs context window, helping users who should quickly process massive pieces of text. Service requires (minimum) a Google One AI Premium subscription of $ 19.99 per month.
The models of the one released in 2024
Deepseek R1
This Chinese model he took Silicon Valley from the storm. Deepseek’s R1 performs well in coding and mathematics, while his open -source nature means that one can run it into the country. Plus, it’s free. However, R1 integrates the censorship of the Chinese government and faces increasing prohibitions to send user data back to China.
Deep research of Gemini
In -depth research summarizes Google search results in a simple and well cited document. Service is useful for students and anyone else who needs a quick summary of research. However, its quality is almost as good as a current letter revised by colleagues. Deep research requires a $ 19.99 google AI premium reconciliation.
Meta Llama 3.3 7b
This is the newest and most advanced version of Meta Llama open source models. Meta has followed this version as the cheapest and more efficient, especially for mathematics, general knowledge and the following guidelines. Is free and open source.
Openai Sora
Sora is a model that creates realistic text -based video. While it can generate whole scenes than just clips, Openai admits that it often generates “unrealistic physics”. It is currently only available in paid versions of the chatgpt, starting with plus which is $ 20 a month.
Alibaba Qwen Qwq-32b Preview
This model is one of the few that rivals O1 O1 to some industry standards, excellent in mathematics and coding. Ironically for a ‘reasoning model’, it has “the space for improvement in the reasoning of the common sense,” says Alibaba. It also involves censorship of the Chinese government, Techcrunch’s testing says. Is free and open source.
The use of anthropic computer
The use of Claude’s computer aims to take control of your computer to complete tasks like encryption or booking a plane ticket, making it an Openai operator’s ancestor. Computer use, however, remains in beta. Price is through API: $ 0.80 per million signs of data, and $ 4 per million signs of production.
X.Ai’s Grok 2
X.Asi, the company he owned by Elon Musk, has begun an extended version of her flag Grok 2 chatbot that claims to be “three times faster”. Free users are limited to 10 questions every two hours in Grok, while subscribers on Premium and Premium+ Premium+ Planet enjoy higher use limits. X. He also launched an image generator, Aurora, which produces very photorealist images, including some graphic or violent content.
Openai O1
The O1 O1 family aims to produce better answers by “thinking” through responses through a hidden feature of reasoning. The model excels in coding, mathematics and security, Openai claims, but there are also issues that deceive people as well. O1 requires reconciliation in Chatgt Plus, which is $ 20 a month.
Anthropik’s Claude Sonnet 3.5
Claude Sonnet 3.5 is a model claims anthropic as the best in the classroom. Becomes made known for his coding skills and is considered a chosen choice of an insider of technique. The model can be reached free of charge in Claude although heavy users will need a monthly reconciliation of $ 20 pro. While he can understand the images, he cannot generate them.
Openai GPT 4O-MINI
Openai has protected GPT 4o-mine as its most affordable and faster model thanks to its small size. It is intended to enable a wide range of tasks such as empowering chatbots for customer service. The model is available at the free level of chatgpt. Better is best suited for simple tasks with high volume than more complex ones.
Time command R+
The Cohe R+ command model excels in complex use generation (or Rag) applications for enterprises. This means that it can find and quote specific parts of the information really well. (Rag’s inventor actually works in chere.) Still, Rag does not completely solve the problem of her hallucination. Cheere models are for enterprise users.
