AI models are being cracked at a speedy crackdown by everyone from large tech companies like Google and startups like Open A and Entropic. Keeping track of the latest people can be very high.
Increasing confusion is that AI models are often promoted on the basis of industry benchmark. But it Technical measurements often show less About how real people and companies use them.
To eliminate the noise, Tech Crunch has set up a review of the latest AI models released since 2024, with details about their use and for whom they are. We will also keep this list fresh with the latest launches.
There are literally over a million AI models: hugs face, for example, Hosted more than 1.4 million. So this list can lose some models that perform better in a way or in some other way.
AI models released in 2025
Google Gemini 2.5
Gemini 2.5 Pro experimental, a reasoning model, takes the lead in AT Web apps and code agents build up According to Google. However, it performs less on a famous coding benchmark than Claude Swant 3.7. The model requires $ 20 monthly Gemini Advance subscription.
Chat GPT -4 and Image Generator
Open I Has upgraded your current GPT4O model Not just the text, to produce the pictures. Soupup model soon went viral Changing photos to Studio Unseen Style Mobile PhoneDespite the clear copyright concerns. To access GPT-4O, a minimum, $ 20 per month requires subscription besides Chat GPT.
Stable Virtual Camera of Stability AI
Image Generation Startup Stability AI What is a model launch The company says 3D scenes and camera angles from the same 2D image. However, it is still struggling with scenes, including complex elements like humans and moving water. This model is available for the use of non -commercial research on hugging.
Kohir’s vision
Harmony Issued a multi -modal model The Aya vision is called that it claims that it is best to work like answering photos in class and answering questions. Unlike other models, it is mastered in other languages other than English. This is available Free on WhatsApp.
Openai’s GPT 4.5 “Orion”
Open AI made his call The biggest model todayTo disrupt its strong “global knowledge” and “emotional intelligence”. However, it performs less on certain standards than new arguments models. Area is available for Openai’s 200-month project users.
Claude Swant 3.7
Anthropic says it is The industry’s first “hybrid” reasoning modelBecause both of these can eliminate quick answers and think about things when needed. It also gives users control how long the model can think of, Per Entropic. Swant 3.7 is available for all cloud users, but heavy users will need a Month 20-month pro-project.
Zia’s Gok 3
Grok is 3 The latest flagship model Startups made from Elon Kasturi. It is The claims Improve other leading models related to mathematics, science, and coding. After a study the model needs X premium (which is $ 50 per month) Found Grook 2 bent to the left, musk Promised To move Grook more “politically neutral” but it is not yet clear whether it has been achieved.
Openi O3-mini
It’s Openi The latest reasoning model And has been improved for stem -related tasks such as coding, math and science. It is Not the most powerful of the Openi Model but because it’s small, company Says This is significantly low cost. It is available for free but requires subscription for heavy users.
Openi deep research
Openi has a deep research Is designed to do deeper research On a topic with clear references. This service is only available with Chat GPT Month 200 per month Pro Subscript. Open I It recommends For everything from science to shopping research, but beyond it The deception has become a problem For AI.
Wrong to the cat
There is a misunderstanding Launted the app version of Le ChatA multi -modal AI personal assistant. False Claim Lee Chat responds faster than any chat boot. It also has a paid version The latest journalism From AFP Test from the world Lee Chat’s performance was impressive, though he made more mistakes with Chat GPT.
Openi operator
Openai’s operator Have to be A personal interns that can work freely, as helps you buy grocery. This requires a month 200-month chat GPT Pro subscription. AI agents have promised a lot, but they are still experimental: Washington Post reviewer The operator says Payed with the reviewer’s credit card, a dozen eggs decided to order a dozen eggs for $ 31.
Google Gemini 2.0 Pro experimental
Google Gemini’s Very waiting flagship model They say that it takes the lead in coding and understanding ordinary knowledge. It also has a very long context window of 2 million tokens, which helps users who need to take a wide range of texts. For service (minimum) a Google One AI Premium Subscription requires a month. 19.99.
AI model released in 2024
DPSEC R1
These Chinese AI model Taken Silicon Valley by storm. DiPsic’s R1 performs well on coding and math, while its open source nature means that anyone can run it locally. Also, it’s free. However, R1 makes integration Chinese government censorship and Face -growing restrictions Possibly send user data back to China.
Gemini deep research
Deep research Google offers a summary of search results In a simple and well -presented document. Service It is helpful for students and someone else who needs immediate research summary. However, its quality is not as good as a peer reviewed paper. For deep research .9 19.99 Google One AI Premium Subscription is required.
Round calls 3.3 70b
It is Modern and sophisticated version Open Source Lama AI models of Meta. Meta has defeated This version Since it is still the cheapest and most efficient, especially for mathematics, general knowledge, and instructions. It’s free and open source.
Openi Surah
Surah is a model Creates realistic videos Text -based when it can only create whole scenes instead of clips, Openi confessed That it often produces “non -realistic physics”. It is currently available only on the paid version of Chat GPT, which begins with Plus, which is $ 20 a month.
Alibaba qwen qwq-32b-preview
This model is One of a few for rivals By performing well in mathematics and coding, some industry benchmarks Open O1. The irony is that for the “model of reasoning”, it has the potential to improve the reasoning of the intellect, ” Alibaba says. This includes censorship of the Chinese government, Takkarch Testing Shows. It’s free and open source.
Anthropic’s computer use
Claude’s computer use means Control your computer IT to complete the tasks such as coding or booking aircraft tickets, making it an open operator. The use of a computer, however, The son lives in. Pricing is from API: Input 7 0.80 per million tokens and $ 4 per million token output.
Zia’s Gok 2
Elon Musk’s AI Company, Z, has launched a launch Better version of his flagship Grook 2 chat boot These Claim “Three times faster.” Free users are limited to 10 questions on GROK every two hours, while users of X’s premium and premium+ projects enjoy the extent of use. Zee also launched an image generator, aurora Produces highly photographed photosIncluding some graphic or violent material.
Open AO 1
Openai’s O1 Family Is to prepare better answers through “thinking” through the hidden answers Feature of reasoning. Model takes the lead in coding, math and safety, Openi claimsBut is Problems trying to deceive humans too. Chat GPT Plus needs to subscribe to the use of O1, which is $ 20 a month.
Anthropic’s Claude Sant 3.5
Claude Sant 3.5 is a model anthropic Claims to be the best in class. It is famous for its coding abilities and is considered tech internal Choosing chat boot. The model can be accessed free of charge on the cloud, though heavy users will need $ 20 monthly pro -subscriptions. Although it can understand the images, it cannot produce them.
Openi GPT 4O-MINI
Open is Toted GPT4O-Mini Thanks to its smaller size as its cheapest and fastest model. That means Customer Service Chat to enable widespread tasks such as strengthening boats. Model Chat is available at the free level of GPT. It is better suitable for easy -to -work tasks than more complex tasks.
Cohere Command r+
Well Command r+ model Excel in a growing generation (or RAG) requests for complex recovery for businesses. This means that it can really find and refer to specific pieces of information. (The inventor of the rig Actually works in Kohar.) Still, rags AI does not fully solve the fraudulent problem.