Subora

General Reasoning Model

True grounding, no memorization

Ask for exposure

Energy to reach general intelligence

Human (16 yrs)

10 GJ

Frontier AI+

~100 PJ

10,000,000× more energy. Still can't reason.

ARC-AGI-2 — Before memorization (March 2025)

GPT-4o

~0%

Claude 3.7

0.9%

DeepSeek R1

~1%

o3 ($20k/task)

3-4%

Human ($17/task)

60%

Novel benchmark. Models collapse.

ARC-AGI-2 — After memorization (Dec 2025)

GPT-5 Pro ($7/task)

18.3%

Opus 4.5 ($2/task)

37.6%

Gemini 3 ($77/task)

45.1%

Human ($17/task)

60%

Human collective

100%

Dataset leaked. Scores climb. Still not reasoning. Source

ARC-AGI-3 drops in 2026. Back to 0%.

The cycle repeats until the architecture changes.

Day 45.