Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

ARC-AGI

ARC-AGI is a series of benchmarks, started by François Chollet and run by ARC Prize, that tests whether an AI system can solve new reasoning puzzles it has not seen before, measuring how efficiently it learns rather than what it knows.

Also called: Abstraction and Reasoning Corpus, ARC-AGI-2, ARC-AGI-3, ARC Prize

By Divya Sharma · Checked Nov 15, 2026 · How we review

Most benchmarks test skill at a task. Chollet's 2019 paper On the Measure of Intelligence argued that skill alone is a poor measure, because unlimited training data lets developers buy skill without real generalisation. He proposed defining intelligence as skill-acquisition efficiency and released the Abstraction and Reasoning Corpus (ARC) to test it.

How it is designed

ARC tasks rely only on core knowledge priors: cognitive building blocks that people have from birth or pick up very early. They avoid cultural or specialist knowledge, so humans and AI can be compared fairly. ARC Prize says the benchmark targets fluid intelligence, the ability to reason and adapt to new problems, rather than crystallized intelligence built from accumulated knowledge. Each task is new, so memorising the training data does not help much.

The series

The original set is now called ARC-AGI-1. ARC Prize has since released ARC-AGI-2 and ARC-AGI-3 and runs yearly competitions, each with a public leaderboard. Check the ARC Prize site for current scores, as they change with each version.

Why it matters

ARC-AGI is designed to resist memorisation, which makes it a common reference point for reasoning models and test-time compute. Its definition of AGI, matching the learning efficiency of humans, is one view among several. A high score shows adaptability on these puzzles, not general competence, so read it alongside other benchmarks.

Related terms

AI benchmark
An AI benchmark is a fixed, public set of tasks with an agreed scoring method, used to compare models on the same test, for example coding issues solved, exam questions answered or retrieval quality across many datasets.
Reasoning model
A reasoning model is a language model trained to work through a problem in a stream of intermediate thinking tokens before giving its final answer, which tends to help on maths, coding and multi-step tasks.
Test-time compute
Test-time compute is the amount of computation a model spends while answering a request, rather than during training. Letting a model think longer or try several answers can improve results on hard problems.
Leaderboard (AI)
An AI leaderboard is a public ranking of models by their scores on one or more benchmarks or on human preference votes, updated as new models are tested, so readers can compare models side by side.

Sources