The Pulse
ARC Prize Sets ARC-AGI-4 on Autonomous Invention
ARC Prize says ARC-AGI-4 will target autonomous, open-ended innovation, while leaving the benchmark’s design unspecified.
AI.info Team ·
ARC Prize Moves Toward Open-Ended Innovation
ARC Prize has announced the direction for ARC-AGI-4, describing the planned benchmark as a test of autonomous, open-ended innovation. The statement identifies invention as the capability the organization wants its next benchmark to address, while leaving the benchmark’s specific design undisclosed.
“ARC-AGI-4 will be a benchmark for autonomous open-ended innovation.”
ARC Prize, nonprofit AI benchmark organization
In its announcement, ARC Prize said the benchmark will continue the organization’s commitment to open source and provide the research community with a shared target for progress. It argued that, despite progress in AI models, humans still significantly outperform AI at open-ended invention.
ARC Prize characterized invention as a meta-skill that can drive advances across fields of technology. It said advanced AI capable of scientific innovation could produce new technology, knowledge and understanding, and presented that outcome as a positive-sum future.
What the Announcement Does—and Does Not—Specify
The announcement establishes an intended research direction rather than a finished evaluation. ARC Prize did not provide tasks, scoring rules, submission requirements, benchmark environments or an evaluation protocol for ARC-AGI-4.
That leaves unanswered how ARC Prize will define an invention for benchmark purposes, how it will distinguish useful work from unusual but unproductive output, and how results will be assessed. The announcement also does not specify whether ARC-AGI-4 will be based on software environments, scientific problems, physical experimentation or another format.
For now, the statement is a commitment to the type of capability ARC Prize intends to measure, not a release of a test that researchers can run or enter.
From Static Reasoning to Interactive Environments
The ARC benchmark series has changed format as it has pursued harder forms of generalization. ARC-AGI-1 and ARC-AGI-2 use abstract reasoning tasks in which systems infer transformations from examples and apply them to new inputs. ARC-AGI-3 is an interactive reasoning benchmark built around environments in which agents must adapt to unfamiliar situations.
ARC Prize’s ARC-AGI-3 materials describe its toolkit as an interface for interactive environments, allowing agents to take actions and receive observations. The 2026 competition describes ARC-AGI-3 as an interactive reasoning benchmark for AI agents, alongside a separate ARC-AGI-2 static-reasoning track.
ARC-AGI-4, as described in the announcement, would shift attention from solving or acting within a supplied setting toward autonomous innovation. But ARC Prize has not yet said how that shift will be implemented or measured.
Open Source as Part of the Proposal
The announcement also makes openness part of ARC Prize’s case for the project. ARC Prize said the knowledge behind frontier AI, as well as the technology itself, should be broadly distributed among researchers, academics and organizations.
It warned that coordinated efforts by the AI industry to reduce openness or concentrate access to frontier AI would undermine the future it described. ARC Prize said it wants a future in which more people can contribute to and benefit from AI progress.
That position is consistent with the organization’s existing competition rules. ARC Prize’s 2026 competition requires participants to open-source their solutions to be eligible for prizes, and says its mission is to accelerate progress toward open AGI by making methods available to the research community.
ARC Prize has set out a broad goal for ARC-AGI-4: evaluating autonomous, open-ended innovation. The organization has not yet released the practical details needed to show how the benchmark will test that goal.