ARC Prize verifies Google Gemini 4 Argon at 97.5% on ARC-AGI-3
Overview
ARC Prize reports that Google's Gemini 4 Argon scores 97.5% on ARC-AGI-3 using a provider adapter harness at $4.8K per run, and 49.8% at $5.8K with the standard harness.
The same post lists 91.3% on ARC-AGI-2 at $0.32 per task and 98.5% on ARC-AGI-1 at $0.16 per task. ARC Prize says Google DeepMind is the second frontier lab to score above 95% on ARC-AGI-3.
ARC Prize says Argon used 62% fewer actions than the human baseline across the 25 ARC-AGI-3 public games, and matched or beat the speedrun record on 3 of them. A separate ARC Prize post says OpenAI's GPT-6 Astra scored 99.9% on the same benchmark, placing Argon second. Mike Knoop says Argon is only the second model family to officially max out ARC-AGI-3 using a provider adapter. François Chollet says Argon's performance profile is similar to Astra's.
ARC Prize has published the full Argon results on its public leaderboard, along with its open-source arc-agi-benchmarking code and testing policy.
Written by AI from the articles below · updated Oct 15, 1:10 PM ET
Check the sources:
Developments
3 developments
- Oct 15, 12:22 PM ET · 1 articleARC Prize publishes full Gemini 4 Argon benchmark resultsARC Prize: ARC Prize publishes full Gemini 4 Argon results on its leaderboard
- Oct 15, 12:22 PM ET · 1 articleARC Prize reports Google's Argon scores 97.5% on ARC-AGI-3ARC Prize: Google Argon scores 97.5% on ARC-AGI-3, behind GPT-6 Astra
- Oct 15, 12:22 PM ET · 3 articlesGemini 4 Argon scores on ARC-AGI-1, 2 and 3 as verified by ARC PrizeARC Prize: ARC Prize verifies Gemini 4 Argon scores above 95% on ARC-AGI-3
Article timeline
The articles in this story. Times are ET.
François Chollet@fcholletXPickChollet says Google's Argon matches Astra on ARC-AGI-3AIFrançois Chollet says Argon has a performance profile similar to Astra on ARC-AGI-3, and that Google is back at the frontier. ARC Prize reports Gemini 4 Argon scores 97.5% on ARC-AGI-3 with a provider adapter harness at $4.8K, and 49.8% at $5.8K with the standard harness. Google DeepMind is the second frontier lab to score above 95% on ARC-AGI-3.
Mike Knoop@mikeknoopXGoogle DeepMind's Gemini 4 Argon scores 97.5% on ARC-AGI-3 with a provider adapterAIMike Knoop says Gemini 4 Argon is only the second model family to officially max out ARC-AGI-3 using a provider adapter. ARC Prize's verified results show 97.5% at $4.8K with the provider adapter harness and 49.8% at $5.8K with the standard harness. The same results list 91.3% on ARC-AGI-2 at $0.32 per task and 98.5% on ARC-AGI-1 at $0.16 per task.
ARC Prize@arcprizeOfficialARC Prize publishes full Gemini 4 Argon results on its leaderboardAIARC Prize has published full Gemini 4 Argon results on its public leaderboard, along with a reproducible benchmarking repository and its testing policy. The post links to the leaderboard, the open-source arc-agi-benchmarking code, and the full Gemini 4 Argon results page.
ARC Prize@arcprizeOfficialGoogle Argon scores 97.5% on ARC-AGI-3, behind GPT-6 AstraAIGoogle's Argon scores 97.5% on ARC-AGI-3, second only to GPT-6 Astra's 99.9%, according to ARC Prize. Like Astra, Argon uses 62% fewer actions than the human baseline across the 25 public games. It matches or beats the speedrun record on 3 games and comes within 10% of it on 6 others.

ARC Prize@arcprizeOfficialPickARC Prize verifies Gemini 4 Argon scores above 95% on ARC-AGI-3AIARC Prize reports verified ARC-AGI results for Google's Gemini 4 Argon. It scores 97.5% on ARC-AGI-3 with a provider adapter harness at $4.8K, and 49.8% at $5.8K with the standard harness. It also reports 91.3% on ARC-AGI-2 at $0.32 per task and 98.5% on ARC-AGI-1 at $0.16 per task. The post says Google DeepMind is the second frontier AI lab to score above 95% on ARC-AGI-3.

Heat trend
Not enough continuous observations to show a trend yet.