ARC Prize has announced ARC-AGI-3 where humans score 100% and no AI model gets over 1%.
- It features many novel game environments built from scratch with no instructions or explicit goals given to participants.
- AI must explore each environment, acquire goals, form hypotheses, develop strategies and adapt using only core knowledge priors.
- Efficiency scoring now measures skill acquisition by comparing the number of actions AI needs to succeed against human baselines.
- Mike Knoop noted the new stateless client and out-of-distribution private set for fairer testing of on-the-fly world modelling and continual learning.
- Francois Chollet explained the benchmark acts as a moving target that tracks the exact gap between what is easy for humans and hard for AI on the path to AGI.
The demonstration set with sharable replays and failure mode analysis is now live for immediate testing.
@scaling01 I have been saying the same thing for years. Initially I was using this exact line ("it's a compass, not a target to hit") about ARC 1, back around 2021-2022, before ChatGPT. This has always been our stance. My bar for AGI has been public and unchanged since 2019. It's AGI when
Keep in mind: ARC-AGI is *not* a final exam that you pass to claim AGI. Including ARC-AGI-3. The benchmarks target the residual gap between what's hard for AI and what's easy for humans. It's meant to be a tool to measure AGI progress and to drive researchers towards the most
Human-level general intelligence is achieved when an AI system can approach a new task and figure it out, without human intervention, *with the same learning efficiency as humans*. If every new task requires human intervention, it's not general. If every new task requires
ARC-AGI-3 and ARC Prize 2026 are now live with $2,000,000 in prizes! As of today, version 3 is the world's only unsaturated agentic intelligence benchmark. Humans score 100% and frontier AI scores ~0%. Play here: https://t.co/nF1MhGvTyz While no single version of ARC is
For those wondering about ARC-AGI-4 timing: it will be released in early 2027. We are aiming for a yearly release schedule for new benchmarks. We are also aiming for each new benchmark to be fully unsaturated upon release, and to target the most important unanswered research
Kol Tregaskes