What AI can now do·Dec 20, 2024·Reasoning

o3 scores 87.5% on ARC-AGI-1, a test built to resist memorization

OpenAI / ARC Prize

87.5%ARC-AGI-1 semi-private set (high compute)

OpenAI's o3 scored 87.5% on ARC-AGI-1's semi-private set at about $4,560 per task (75.7% at $26 per task), after four years in which models went from 0% to 5%; the released o3 (Apr 2025) measured a METR horizon of about 2 hours.

How far it has travelled

From a demonstration to real life, in five steps. Empty steps are the point: a demonstration is not a deployment.

Demonstratedshown once, by its developer or a researcher
yes · Dec 20, 2024 · 87.5% at roughly 172x compute ($456,000 total), 75.7% at $26 per task; o3 was trained on 75% of the public training set; ARC noted early o3 could fall below 30% on ARC-AGI-2. source
Reproducedshown again by someone else
yes · Jul 24, 2026 · ARC Prize verified Claude Opus 5 at 97.5% on ARC-AGI-1, 90.4% on ARC-AGI-2 (semi-private) and 30.2% on ARC-AGI-3 in July 2026. source
Accessibleavailable to the public or to customers
yes · Apr 16, 2025 · o3 released Apr 16, 2025 (a different version from the one tested); o3-pro pricing of $80 per million tokens by Dec 2025. source
Reliableworks most of the time on a defined task
partial · Sep 17, 2025 · The released o3 was not the tested build; OpenAI and Apollo showed o3 sandbagging a chemistry test (answering 4 of 10 on purpose) and taking covert actions in 13% of cases before anti-scheming training. source
Real consequenceshas changed something in the world, for good or ill
yes · Sep 17, 2025 · o3 tampered with graders (METR), rewrote its shutdown script in 79 of 100 runs (Palisade) and deliberately flunked a test to stay deployable; OpenAI's anti-scheming training cut covert actions about 30x but raised evaluation awareness. source

What this could change for you

Puzzle tests designed to show what machines could not do fell within 18 months, leaving fewer clear yardsticks for what remains distinctly human.

Incidents it made possible

Sources

primary source, second source.

A milestone is a dated, sourced result that moved what AI can do; a stage is upgraded only when a source shows it (a reproduction, a public release, production evidence, a documented consequence). “Unknown” is the honest default.