Which AI Model is Most Used?

TALLINN, EE / ACCESS Newswire / December 17, 2025 / November 2025 will probably go down in history as one of the most action-packed months in terms of AI releases, because there's been more changes to the top of LMArena leaderboard than a monthly weather forecast.

That's because OpenAI released GPT-5.1 on November 12th, which became the best model on LMArena, then Google released Gemini 3 Pro on November 18th, which dethroned GPT-5.1, xAI released Grok 4.1 on November 25th, which was the best model for a short space, but Anthropic released Claude Opus 4.5 on November 30th, which defeated all of them. Shortly after, DeepSeek released V3.2, for good measure.

But while all of that was going on, a bigger story started unfolding over the last couple of days about whether Gemini 3 Pro would finally dethrone ChatGPT as the go-to AI model, because of how good it is.

The talk got loud enough that OpenAI is rushing to release GPT-5.2 on December 9th - less than a month after 5.1 - in what's being called a code red response.

With all this activity, people suddenly have more flagship AI models to choose from than Halloween candy. But the real question is: what are people actually using?

We analyzed usage data from November 28 to December 5, 2025 at Overchat AI - a platform that gives users equal access to every major AI model, and looked at a period in the immediate aftermath of all these flagship releases. The data comes from 437,500 users who generated approximately 2.875 million AI interactions.

Here's what we found:

Understanding New Models

Before we look at the data, let's quickly talk about what each of these models are and why it's significant, because all of them are, in some way.

GPT-5.1 is a midlife update by OpenAI that mostly makes conversations feel warmer. This is probably the least disruptive of the releases, which explains the code red situation we mentioned. There are two types: GPT-5.1 Instant is for everyday tasks, and GPT-5.1 Thinking is a reasoning model. The biggest update is the ability to think in context. The model answers faster when it doesn't need to think and is more accurate when it does. The user doesn't need to choose between always thinking or always not thinking.

Gemini 3 Pro is Google's most intelligent model. It scored 1,501 on the LMArena leaderboard. This was the first model in history to break the 1,500 mark. It's really good at understanding different types of media, like text, images, audio, and video. It also did really well on academic reasoning benchmarks (Humanity's Last Exam - 37.5%, GPQA Diamond - 91.9%). Google positioned this as their answer to GPT-5.

Claude Opus 4.5 is Anthropic's most advanced coding model, and it blew all other models on this list out of the water in coding benchmarks. According to the Anthropic team, it also did better than any human in internal tests.

Grok 4.1 is xAI's latest model and scored 1,483 Elo on LMArena, which is really close to Gemini 3 Pro. Grok 4,1 thinking is currently in second place overall.

DeepSeek V3.2 is the latest top model from the Chinese AI lab DeepSeek. It costs 10 times less to run, and it performs better than any other model on this list when it comes to math tests.

Most Used AI Models

We've analyzed approximately 2.875 million interactions with AI chatbots on Overchat AI (including image and video models and tools) and here's the data from November 28 to December 5, 2025:

The takeaway

Despite the ongoing talk about Gemini 3 Pro possibly stealing the crown from OpenAI and ChatGPT, users still prefer GPT-5.1 a lot. It captured 66.0% of usage, which is almost two-thirds of the total. Meanwhile, Gemini 3 Pro came in fourth place with just 8.1%.

Of course, this is only part of the picture - it doesn't reflect what's happening with enterprise adoption and corporate clients, which is a completely different conversation. And OpenAI's rush to release GPT-5.2 is likely driven as much by perception and PR as by performance. They want to reinforce their position as the trendsetters in the space.

But what's clear from the data is that the general public isn't easily swayed by benchmarks or LMArena rankings. They trust the GPT brand. It's the chatbot they know, and it's the one they keep choosing.

Ironically, in the B2C space, GPT-5 seems to be feeling more pressure from an open-source Chinese model than from Google or Anthropic.

Nano Banana Pro Is the Most-Used Image Generation Model

OpenAI may dominate the chatbot space, but it's a different story when it comes to image generation. Google is still the uncontested number one here, as almost nobody chooses OpenAI's GPT-Image-1 model.

On November 22, 2025, Google released Nano Banana Pro along with Gemini 3, and AI enthusiasts immediately said that it was as groundbreaking as Flux Knotext.1 was at the time, when it came out on May 29, 2025 (time really goes fast in the world of AI).

Built on the Gemini 3 Pro foundation, Nano Banana Pro is very realistic, very accurate, and very knowledgeable. It can blend up to 14 images at a time or create infographics with perfect text and accurate facts about anything.

The downside to all that is that this quality comes at the expense of speed and price. The Nano Banana 2 costs as much as some video models, rather than image models, per generation.

Because of that, Google positions it as a premium model that exists alongside the original Nano Banana - hence the Pro in the name.

With that in mind, it's surprising that despite the extra cost people almost stopped using the original Nano Banana:

For example, on Overchat AI Nano Banana Pro costs 100 credits per generation, which makes it 5-10x more expensive than other models.

• Users can generate 80 images per week with Nano Banana Pro on Overchat AI

• Users can generate up to 800 images per week with other models on Overchat AI

The takeaway

Despite this significant price premium, Nano Banana Pro version generated 310,720 chats compared to the original's 29,855 chats. This is in contrast to text models, where we often see users gravitating toward faster, cheaper options for everyday tasks. With images users want the best possible result and don't mind paying the premium to get it.

Sora 2 is The Most Used Video Generation Model

When it comes to video generation, the tug-of-war between Google and OpenAI isn't even close - Sora 2 completely outclasses Veo 3.1 and has become the most viral AI model of 2025.

The table below shows the three most-used AI models, and Veo 3.1 doesn't even make the list. Sora 2 alone accounts for more than 80% of all video generations across major platforms - an overwhelming win for OpenAI.

The takeaway

These results aren't that surprising given that Sora has been the most marketed AI video generator, with OpenAI building significant hype around its capabilities.

Interestingly, Wan 2.5 captures 30.9% of usage, and Seedance 1.0 Pro - 7.3%, which is ahead of Veo 3.1. It's also noteworthy how concentrated video model usage is compared to text models. In text, we saw a 66/34 split between the top two models. In video, it's 85/15. When it comes to video generation, users have a much clearer preference for the established leader.

Here's the updated Bottom Line section:

Bottom Line

Here are the key findings from 437,500 users and 2.875 million AI interactions from November 28 - December 5, 2025:

• Unsurprisingly, Sora 2 is the most widely used video generation model. Over 80% of all videos generated on Overchat AI use Sora 2. Wan and Seedance follow with 30% and 7%, respectively, while Veo 3.1 is overlooked.

This analysis is based on usage data from Overchat AI between November 28 and December 5, 2025, representing 437,500 users and approximately 2.875 million AI interactions.

About Overchat AI

Overchat AI is an independent AI aggregation platform that provides users with equal access to all major artificial intelligence models across text, image, and video generation. By allowing real-world, side-by-side usage of competing models, Overchat AI delivers unique insights into how people actually interact with AI - beyond benchmarks, marketing claims, or hype cycles. Its anonymized usage data is widely used by researchers, developers, and enterprises to understand emerging trends in AI adoption.

Media Contact

Name: Alex Morgan
Title: Head of Research
Company: Overchat AI
Email: press@overchat.ai
Phone: +1-415-555-0198

SOURCE: Overchat AI