What GAIA scores mean — How far have AI agents progressed?
When the GAIA benchmark was first released, even GPT-4, which was the highest-performing model at the time, scored only around 30%. This meant that, given GAIA’s requirements for complex reasoning, tool use and multi-step problem-solving beyond simple question answering, AI was still at an early stage in its ability to “think and act”.
Today, however, leading agents have reached 92.36%.
This change is more than a simple improvement in performance. It indicates that AI has now moved beyond answering questions and entered the ‘Agentic AI’ stage, where it interprets situations, selects the tools it needs, and plans and executes multiple steps autonomously.
In just a few years, AI has evolved from a “knowledge-generation tool” into an “agent that carries out work”.
📌 And CLEVI is among the top 2.5%
In this global competitive landscape, the fact that CLEVI’s domestically developed agent has been listed among the top 2.5% on the GAIA leaderboard demonstrates technological self-reliance and proves that an AI agent made in Korea can meet global standards. This means more than just a number. It signifies technological capability objectively validated through competition with thousands of models on a publicly available global benchmark, rather than in a specific demonstration environment.
More importantly, this achievement was not the accidental result of a single model. It shows that agents with different designs have repeatedly delivered top-tier performance on the same Agent Stack.
In other words, CLEVI’s technology is not a “one-off strong result”, but reproducible structural competitiveness and platform-level capability that can be expanded across various industries and scenarios.
So, what does this metric mean?
From the user’s perspective, the biggest barrier to adopting AI is the question: "Can I really trust it with my work?"
Performance repeatedly proven on the stage of a global public leaderboard—not through flashy demos or company presentations—is the most objective answer to that question. Specifically, it is meaningful in three ways.
First, reducing implementation risk
Connecting unproven AI to internal operations is a significant burden for businesses.
Performance on credible benchmarks such as GAIA can be directly used as a basis for trust during the technical evaluation stage.
Second, making complex task automation a reality
A ranking in the top 2.5% indicates a level of intelligence that goes beyond simple repetitive tasks, understanding context and independently making decisions across multiple steps to complete them.
Areas of work previously considered "too difficult to entrust to AI" can become practical targets for automation.
Third, securing data protection and performance simultaneously
The proprietary Agent Stack architecture means that data flows do not leave the organisation.
This makes it possible to move beyond the existing dilemma of having to compromise security in order to use high-performance AI.
This architecture is a decisive differentiator, particularly in industries with highly sensitive data, such as finance, healthcare and legal services.
The reproducibility of the architecture is already beginning to be demonstrated.
And the performance of each agent built on that architecture will soon lead to tangible change in the field.
"Ultimately, technological maturity is completed by 'confidence' in the field."
CLEVI does not simply provide high-performance AI. Our essence is to build 'execution-focused infrastructure' that breaks down the security barriers businesses face and can actually complete complex practical work.
Now, move beyond considering adoption and start a safe, powerful AI work environment with CLEVI.
As a dependable partner you can trust with your work, we will work with you to create tangible innovation in your business.