AI Trends

ARC Prize evaluation: GPT-6.2 Astra uses 64.1% fewer actions than the human baseline on ARC-AGI-3

ARC PrizeUpdated Nov 18 02:31

AI summary

ARC Prize published the evaluation results for GPT-6.2 Astra on ARC-AGI-3: across 25 public games under the max reasoning setting, its action count was 64.1% lower than the human baseline and 5.2% lower than GPT-6 Astra. On ARC-AGI-2 semi-private tasks, GPT-6.2 Astra used 51% fewer reasoning tokens than its predecessor at the low through xhigh levels, while using 12% more at the max level. In ARC-AGI-3 semi-private games, its first-action reasoning tokens were 53% lower than GPT-6 Astra's.

Why read this

SelectedAI score 78

ARC Prize published a comparison of GPT-6.2 Astra's action counts and reasoning tokens on ARC-AGI-3, which offers a look at the new model's efficiency trade-offs.

Topics

Tags

Body

@OpenAI Across the 25 ARC-AGI-3 public games at max reasoning, GPT-6.2 Astra used 64.1% fewer actions than the human baseline and 5.2% fewer than GPT-6 Astra with the provider adapter harness: https://t.co/l7WxRXb5t6

Source

ARC Prize