Skip to content
GLOBALANCE INTELLIGENCEThe economy ahead.
This Week Index 28/100 Sign in
Index › All indicators
What MovedSeven DimensionsFour PhasesAll Indicators
← All 37 indicators
PHASE 01 · INDICATOR 06 · AI CAPABILITIES

Reasoning test for novel tasks (ARC-AGI-3): best AI model

Whether an AI genuinely solves new reasoning tasks rather than reproducing memorised answers.

63 / 100 Punkte
PROGRESS TOWARDS 2040 · BUILDING · DIRECTION →
START 2020
0
TODAY
62.7
2040 MARK
100

↑ A rising value signals progress. The dots measure the distance travelled from the 2020 starting point towards the 2040 mark, not the raw score. The figure recorded is the result under the ARC Prize Foundation standard test setup (Semi-Private). The same model reaches 99.9% under a vendor-adjusted setup; that figure does not count.

What the number does not say

Not a forecast: 63 points do not mean the 2040 mark will be reached at a particular time. The points measure the distance covered since the 2020 start, not the measurement itself. And a single indicator says little; only together do the 37 form a picture.

Watch this threshold

Save this indicator. The home page then shows how it has moved since you saved it.

Stored only in your browser. No account, no data with us.

CONTEXT

The ARC test measures whether a machine can genuinely learn something new, not merely recall what it has memorised. The third edition (since March 2026) places AI systems in unfamiliar game environments without instructions: they must discover the rules themselves and reach the goal as efficiently as a human would. Humans solve virtually all tasks. The best model manages 62.7 per cent; the second-best reaches 30. The gap between the two is as revealing as the gap to human performance. The earlier tests are saturated: ARC-AGI-1 at 98 per cent, ARC-AGI-2 at 95. For this reason the indicator follows a fixed rule: it always measures the most recent test that machines have not yet mastered, and switches as soon as the best verified score exceeds 85 per cent. Each switch is marked in the chart.

"If the best model approaches the 85% mark on ARC-AGI-3 as well, machines are learning in unfamiliar environments almost as efficiently as humans. That would be the strongest single signal for general AI that exists today."

From the baseline 0 (2020) the indicator is moving towards 100. Current status: 62.7. This corresponds to 63 of 100 Points.

Source: ARC Prize Foundation — ARC-AGI-3 · As of 26 September 2026 · arcprize.org/arc-agi/3

FURTHER INDICATORS IN THIS DIMENSION · AI CAPABILITIES

What comes next