This new benchmark could expose AI’s biggest weakness

The influential AI researcher François Chollet has long argued that the field measures intelligence incorrectly, that popular benchmarks reward a model’s ability to memorize vast amounts of data rather than navigate novel situations and learn new skills. Only recently, with the rise of autonomous AI agents, have companies begun to take that critique seriously. On Tuesday, the ARC Prize Foundation, which Chollet founded with Zapier cofounder Mike Knoop, released a new and more difficult version of its benchmark. The test, called ARC-AGI-3, may offer the clearest measurement yet of how close today’s AI agents are to human-level intelligence.

It consists of more than a thousand simple, video-game-like scenarios designed to measure on-the-fly reasoning rather than memory recall. “You can always achieve skill by memorization by effectively just storing a lookup table of everything you need to do,” Chollet says. “Intelligence is the efficiency with which you’re going to make sense of new things, of new tasks that you’ve never seen before.”

Given no instructions, an agent must develop an understanding of the game environment and its rules, then apply that knowledge to form a strategy across multiple steps toward an ultimate goal. Agents that reach those goals using fewer, more efficient steps earn higher scores, with their creators eligible for all or part of a $1 million prize. As in previous ARC benchmarks, humans can navigate the tasks with relative ease, while many AI systems struggle.

A high score on ARC-AGI-3 could also serve as evidence of artificial general intelligence (AGI). To do “most economically valuable work” performed by humans, as one common definition of AGI requires, AI agents will need to reason through unfamiliar situations in unfamiliar environments. They will need to form abstractions from past experiences and generalize them to new problems they were not explicitly trained to solve.

“I just love that this benchmark basically goes at the heart of this gap that exists between actually measuring for AGI and the standard set of benchmark suites that the big labs and essentially everybody seems to use in the rat race of getting 0.5% of improvement over every other state-of-the-art model for a week,” says Andy Konwinski, whose Laude Institute donated $25,000 to the ARC Prize.

Origins

When the first ARC test was released in 2019, the transformer architecture behind today’s AI chatbots was only two years old, and models were just beginning to generate coherent responses to prompts. Because they could not yet reason in real time, they solved almost none of the ARC-1 puzzles, which limited the benchmark’s adoption.

Chollet saw a fundamental problem with how the industry evaluated progress. Systems that could handle tasks described as “PhD-level” intelligence were failing at simple puzzles. “When the most advanced AI systems are stumped, but a child can do it, that’s a big red flashing light telling you that we’re missing something, that something really important is off,” he says.

The early ARC-AGI-1 results also pointed to a deeper issue with the industry’s strategy for improving its AI: “I think actually ARC is literally the most important unbeaten benchmark in the world because it is the only really clear evidence that contradicts the scaling story that was so dogmatic in the Bay Area in 2023 and 2024,” Knoop says. At the time, the AI labs were confident that continuing to supersize its models, training data, and computing power would continue yielding intelligence gains and eventually lead to AGI. But those systems remained static at inference time (while interacting with a user) and relied only on the pre-trained model weights to generate answers.

Scaling to reasoning

That began to change in 2024, as AI labs started focusing on autonomous agents and the real-world work they might perform. “Deep learning models got to the point where they had accumulated so much knowledge that you could start building a reasoning layer on top of them,” Chollet says. A shift was underway. New reasoning models, such as OpenAI’s o1, released in September 2024 as a research preview, could break complex tasks into smaller parts and evaluate multiple pathways to a solution.

“It was finally trying to address the problem of fluid intelligence, which was missing from the deep learning paradigm,” Chollet says. Researchers began paying closer attention to ARC because it was designed to capture that capability. “[ARC] became this very high signal reference point,” he says. The o1 model improved on earlier results, scoring 21% on ARC-AGI-1, compared to 9% for GPT-4o, its predecessor.

It wasn’t until the OpenAI o3 model, released in January 2025, that new reasoning capabilities significantly affected ARC scores. The model scored between 75% and 87%, depending on the amount of compute used, approaching human-level performance.

Those gains suggested the ARC benchmark might soon be oversaturated. As more models began scoring highly, questions emerged about whether those results reflected genuine reasoning or optimization for the benchmark itself. AI labs were already using engineering workarounds and specialized systems to boost performance. In May 2025, the ARC Prize Foundation introduced ARC-AGI-2 to make the test more resistant to those tactics.

The o3 model that had scored roughly 87% on ARC-AGI-1 initially dropped to just 3–4% on ARC-AGI-2.

Improvement or “benchmaxxing”?

Labs continued finding ways to improve their ARC scores. They began creating specialized software “harnesses” that orchestrated repeated reasoning attempts, then evaluated them and iteratively improved on them. Researchers debated whether the software harnesses reflect the kind of fluid reasoning ARC is meant to measure.

Chollet believes OpenAI spent “tens of millions” on compute in 2025 to train models specifically for ARC-AGI-2, using publicly available ARC puzzle samples to generate additional training data. “What this amounts to is preemptive brute forcing … by trying to guess in advance every possible task,” he says.

At any rate, the tactics worked: top scores rose to 40–50% by December 2025, Knoop says.

“I expect the same will happen with ARC-3, but with ARC-3 it’s going to be harder,” Chollet says. “It’ll be more expensive.”

ARC-AGI-3 arrives at a pivotal moment, as companies and investors bet trillions that AI agents will take on large portions of knowledge work. Models are improving quickly, but they may still lack the intuition needed to handle the complexity and uncertainty of real-world tasks. Anything less risks falling short of true autonomy.

Agents will likely get a grace period in which human workers train and correct them. After that, they will need to build trust and expand their responsibilities. If they fail, businesses may hesitate to adopt them more broadly.

Are today’s agents good enough to earn that trust? If not, how will we know when they are? ARC-AGI-3 could help answer those questions.

It’s a good sign that the AI labs are paying close attention. “I have felt much more pull from the frontier labs and excitement about version three [ARC-AGI-3] than I ever felt about one and two,” Knoop says. The AI labs will work to drive their models toward higher scores on the new benchmark during 2026, and in doing so they may become even more focused on building the qualities and capabilities agents will need to do real-world work. “I think this is just some recognition from a lot of the frontier labs that we do need new ideas,” Knoop says. “We have not figured it all out.”

View the full article