The ARC Prize Foundation has officially announced the upcoming launch of ARC-AGI-4. Founded by AI pioneer François Chollet, the initiative is reacting directly to the unexpectedly swift progress of frontier models. Rather than relying on isolated reasoning puzzles, the new benchmark suite aims to measure the ability of artificial systems to achieve autonomous, open-ended invention in complex problem spaces.
The transition became necessary after the previous evaluation suite, ARC-AGI-3, was saturated significantly faster by frontier models than researchers had previously forecasted. Static evaluation frameworks quickly lost their ability to differentiate genuine general intelligence from advanced pattern recognition. The foundation concluded that static benchmarks could no longer serve as a reliable yardstick for true cognitive adaptability.
ARC-AGI-4 fundamentally pivots toward autonomous innovation and the generation of novel scientific hypotheses. Under the new protocol, artificial agents must operate in open-ended environments to produce entirely new artifacts without relying on pre-existing training sets. This methodology aims to test whether an artificial intelligence can formulate scientific insights and discover solutions completely independently.
To support the development of this next-generation benchmark, the ARC Prize Foundation secured substantial outside backing. The organisation received a dedicated grant of 1,001,337 US dollars from General Intuition. These funds are earmarked directly for task design and infrastructure, ensuring the initiative remains fully independent of commercial cloud and lab sponsors.
Beyond the technical evaluation criteria, the announcement carried a sharp strategic message directed at policymakers and tech conglomerates. The foundation reiterated its strong commitment to open-source software and open research. At the same time, it issued a clear warning that coordinated slowing of frontier AI progress, widely discussed as pacing, threatens to foster an unhealthy cartelization of research among a handful of tech giants.
The unveiling of ARC-AGI-4 marks a pivotal moment for both artificial intelligence engineering and AI governance. It challenges labs to move past brute-force memorization toward systems capable of genuine creative problem-solving. Furthermore, by defending open research against regulatory enclosure, the initiative provides a vital independent counterweight in the escalating debate over frontier model deployment.