Arena Founder Anastasios Angelopoulos on AI Trends for 2026

Because coding happens to be one of the most economically valuable use cases of AI, we see the gap closing faster and faster between open and closed source.

Well, there's so much investment in AI these days. We should expect to see increased diversity and fragmentation. By that I mean, gone are the days where one number can describe how good a model is. There's gonna be so many models. Every frontier lab is going to have several releases in 2026, major ones. And that's gonna be across small models and big models. It's gonna be domain-specific models like coding models and more. And then there's, in addition to those four or five names that everyone knows, there's gonna be dozens of different smaller labs, especially in multimodal, that are developing amazing models with different tastes and different strengths and weaknesses. But it also means that users are gonna be increasingly dealing with the fragmented market, which means that they're going to need places to look to say, hey, what even are the new releases? How can I deploy them in a unified way to my infrastructure? Cerebras has a big place in that in open source. We expect to see American open source models this year as well coming out, which is very exciting.

So LM Arena works differently from a standard benchmark and is more geared around measuring the real world utility of AI. And let me just explain that by first drawing a distinction between the standard static benchmark and how LM Arena works. Static benchmarks, what they are is essentially a test. It's a set of questions in a domain specific vertical that somebody goes out and collects. Let's take a hundred multiple choice questions about ancient history, and we're gonna give them a model and see how well it performs. The issue is that when you keep testing models over and over and over again on the same test, they end up memorizing the test, but they might not actually know anything about ancient history. LM Arena works differently in the sense that we don't have a fixed set of questions. Actually, all the questions that we get every day are new. We've got tens of millions of users every month that are coming to LM Arena and asking questions about whatever they want. And so they end up being very diverse. It can be anything from math to coding questions, to questions about business, to questions about law, to medicine, to romance. You know, anything that people are asking about in the real world, they're also asking about on LM Arena. It's a very global community, hundreds of languages. By virtue of this diversity, the benchmark that you get reflects a lot of unexpected behaviors that you get when you deploy a model in the wild. And by virtue of that fact, each model only sees any question once. And so it can't overfit them. So that's structurally what makes our benchmark different. And of course, there's always a place for static benchmarks to analyze very specific, you know, domain-specific behavior. But LM Arena, you can think about as the sort of source of ground truth for how they operate in the wild.

Well, our philosophy on LM Arena is to source organic usage data from what we call "unhirable experts". People that never would be the ones to write these questions on benchmarks because they're actually busy doing their job. They have a calling. We can actually find them organically within our data because they love using LM Arena. We actually released a blog post on this recently called Arena Expert. We have 5% to 8% of our users end up being these PhD-quality, really high-taste users. And we identify them by looking at their prompt history via like LLM judges. And so once we know who those users are, then we can segment them also by occupational category like legal, medicine, business, marketing and so on, then we can construct a leaderboard that says, hey, if you really care about what those super high-taste, very niche professionals are doing, LM Arena is the place to go.

The most valuable thing about LM Arena is that we're exposing models to real users. And real users are doing things on our platform with AI. But that doesn't mean that they will be moving forward. Or maybe they will be for some aspects, but not others. For example, why do we necessarily need humans to be grading responses for factuality? Humans get facts wrong all the time. I believe that moving forward, we're going to have AI judges that are better at judging factuality than humans. And so there's going to be some aspects that only humans will be able to do. Things like taste. Things like how well is the model communicating. It's almost like an HCI question. How well the model communicates with me. And I imagine us evolving more in that direction where we have a hybrid approach.