The AI model is cranked out at an eye-opening pace, from big companies like Google to startups like Openai and humanity. Tracking the latest is overwhelming.
In addition to the confusion, AI models are often promoted based on industry benchmarks. However, these technical indicators reveal little about how people and businesses actually use them.
To get through the noise, TechCrunch has compiled an overview of cutting-edge AI models released since 2024. We’ve recorded details about how they’re used and what they’re best suited to. This list will also be updated with the latest releases.
There are literally more than a million AI models. For example, Huggingface hosts over 1.4 million people. Therefore, this list may miss models that somehow improve performance.
AI model released in 2025
Openai O3-Mini
This is OpenAI’s latest inference model, optimized for STEM related tasks such as coding, mathematics, and science. Although it is not Openai’s most powerful model, its small size means that the company says it costs significantly lower. It is free to use, but requires a heavy user subscription.
Openai Deep Research
Openai’s deep research is designed to provide in-depth research on topics with clear citations. This service is available only with ChATGPT’s $200 per month Pro subscription. Openai recommends it for everything from science to shopping research, but it should be noted that hallucinations remain an AI problem.
Mistral Le Chat
Mistral has launched the APP version of Le Chat, a multimodal AI personal assistant. Mistral claims that Le Chat responds faster than any other chatbot. There is also a paid version with AFP’s latest journalism. In Le Monde’s tests, Le Chat’s performance was impressive, but it caused more errors than ChatGpt.
Openai Operator
The Openai operator is intended to be a personal intern who can do things independently to help you buy groceries. You will need a Chatgpt Pro subscription of $200 per month. AI agents have many promises, but they are still experimental. A Washington Post reviewer says the operator has decided on his own to order a dozen eggs for $31 paid by the reviewer’s credit card.
Google Gemini 2.0 Pro Experimental
Google Gemini’s long-awaited flagship model states it is excellent at coding and understanding general knowledge. It also has a very long context window of 2 million tokens, which is useful for users who need to quickly process large chunks of text. This service requires (at least) $19.99 a month for a Google One AI Premium subscription.
AI model released in 2024
Deepseek R1
This Chinese AI model took Silicon Valley by storm. Deepseek’s R1 works well in coding and mathematics, but the open source nature means that anyone can run locally. Plus, it’s free. However, R1 faces an upward ban as it consolidates Chinese government censorship and could send user data back to China.
Gemini Deep Research
Deep Research brings together Google search results into simple, well-cited documents. This service is useful for students and others who need a brief research summary. However, its quality is not as good as actual peer-reviewed papers. Deep Research requires a $19.99 Google One AI Premium subscription.
Metalama 3.3 7b
This is the latest and most advanced version of Meta’s open source Llama AI model. Meta promotes this version as the cheapest and most efficient, especially for mathematics, general knowledge, and the following instructions. It’s free and open source.
Openai Sora
SORA is a model that creates realistic videos based on text. While you can generate the entire scene rather than just a clip, Openai admits that it often produces “unrealistic physics.” Currently only available in the paid version of CHATGPT. This starts with Plus, which is $20 a month.
Alibaba QWen QWQ-32B-PREVIEW
This model is one of the few models comparable to Openai’s O1 in benchmarks in a particular industry, and is excellent at mathematics and coding. Ironically, with regard to the “inference model,” Alibaba says, “there is “a room for improvement in common sense reasoning.” It also incorporates a test show from the Chinese government’s censorship, TechCrunch. It’s free and open source.
Humanity’s computer use
Claude’s computer usage is intended to control the computer to complete tasks such as coding and booking plane tickets, and to become the predecessor of Openai’s operators. However, computer usage remains in beta. The price is via the API and is $0.80 per 100 tokens for input and $4 per 100 tokens for output.
X.ai’s Grok 2
X.ai, an AI company owned by Elon Musk, has announced an expanded version of its flagship Grok 2 chatbot, claiming it is “three times faster.” Free users are limited to 10 questions every two hours on Grok, while X’s Premium and Premium+ plan subscribers enjoy higher usage restrictions. X.AI also launched Aurora, an image generator that generates highly lit images with graphics and violent content.
Openai O1
Openai’s O1 family aims to produce better answers by “thinking” through responses through hidden inference functions. Although Openai argues that models are excellent in coding, mathematics and safety, there are also problems that deceive humans. O1 requires you to subscribe to ChatGpt Plus, which costs $20 per month.
Claude Sonnet of Mankind 3.5
Claude Sonnet 3.5 is the best-in-class model human claim. It is now known for its coding feature and is considered by Tech Insider as the chatbot of your choice. Heavy users require a $20 monthly Pro subscription, but models are free to access with Claude. I understand the image, but I can’t generate it.
Openai gpt 4o-mini
Openai advertises the GPT 4o-Mini as the most affordable and fastest model, thanks to its small size. It aims to enable a wide range of tasks, including enhancing the power of customer service chatbots. This model is available in the free tier of ChatGpt. It is suitable for a large number of simple tasks compared to more complex tasks.
Cohere Command R+
Cohere’s Command R+ model excels in complex, searched generation (or RAG) applications for enterprises. This means you can find and quote certain information very well. (The inventor of RAG actually works with Cohere.) Even so, Rag doesn’t completely solve the problem of AI hallucination.
Source link
