ARC Prize compares Google's Carbon and Argon on ARC-AGI-3 public games
Article timeline
The articles in this story. Times are ET.
ARC PrizeOfficialCarbon uses 2% fewer actions than Argon on 23 ARC-AGI-3 gamesAIOn the 23 ARC-AGI-3 public games, both Carbon and Argon won at high with the provider adapter, according to ARC Prize. Carbon used 5,731 actions compared with Argon's 5,846, or 2% fewer. Human speedruns totaled 4,804 actions, so Carbon used 19.3% more than the human reference.
- ARC PrizeOfficialARC Prize reports Carbon used more reasoning tokens than Argon on 24 of 25 games
AIARC Prize says Google's Carbon used more first-action reasoning tokens than Argon on 24 of 25 ARC-AGI-3 public games. The comparison is posted on X by ARC Prize, and the post gives no further detail on the models or methodology.
- ARC PrizeOfficialARC Prize reports Carbon used more reasoning tokens than Argon
AIARC Prize says Carbon used more reasoning tokens than Argon on each of its benchmarks. On ARC-AGI-3 Public games, Carbon at high reasoning averaged 6k reasoning tokens for its first action, compared with 2k for Argon.
