BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

ARC-AGI-3

Semi-private ARC-AGI-3 interactive evaluation set.

As of 24 Oct 2026, GPT-6 Astra leads ARC-AGI-3 on BenchLeader with 99.9%, ahead of GPT-6 Astra at 98.6%, across 69 model configurations with a published result.

Published by
ARC Prize
Category
Reasoning
Index weight
0.5
Models
69
Data as of
24 Oct 2026
Newest result
13 Oct 2026

ARC Prize Foundation, arcprize.org.

What the test looks like

ARC-AGI-3 moves to interactive games: the agent must explore an environment, discover its rules and reach goals without instructions, measuring learning efficiency over many steps.

How it is scored

Percent of the semi-private games completed, run by the ARC Prize Foundation.

What to keep in mind

New, small and unlike the earlier grid tasks; agent scaffolding matters and few models have results.

Top 15
  1. 1GPT-6 Astra (high)99.9%
  2. 2GPT-6 Astra (max)98.6%
  3. 3GPT-6 Astra (xhigh)98.4%
  4. 4GPT-6 Astra (medium)98.4%
  5. 5GPT-6 Astra (low)98.0%
  6. 6Gemini 4 Argon (low)97.5%
  7. 7Gemini 4 Argon (high)97.3%
  8. 8Gemini 4 Argon (medium)97.3%
  9. 9GPT-6 Astra (none)96.7%
  10. 10GPT-6.1 Sol (xhigh)96.4%
  11. 11GPT-6.1 Sol (max)96.2%
  12. 12GPT-6.1 Sol (high)95.0%
  13. 13GPT-6.1 Sol (medium)91.0%
  14. 14GPT-6.1 Sol (low)82.8%
  15. 15Claude Sonnet 5.5 (xhigh)72.4%
69 of 69
#
1GPT-6 Astrahigh99.9%69.62026-09-02
2GPT-6 Astramax98.6%70.42026-09-02
3GPT-6 Astraxhigh98.4%68.92026-09-02
4GPT-6 Astramedium98.4%67.52026-09-02
5GPT-6 Astralow98.0%65.92026-09-02
6Gemini 4 Argonlow97.5%–2026-10-13
7Gemini 4 Argonhigh97.3%70.82026-10-13
8Gemini 4 Argonmedium97.3%–2026-10-13
9GPT-6 Astrano reasoning96.7%59.02026-09-02
10GPT-6.1 Solxhigh96.4%67.72026-09-29
11GPT-6.1 Solmax96.2%69.92026-09-29
12GPT-6.1 Solhigh95.0%67.72026-09-29
13GPT-6.1 Solmedium91.0%67.22026-09-29
14GPT-6.1 Sollow82.8%64.92026-09-29
15Claude Sonnet 5.5xhigh72.4%67.92026-09-28
16Claude Sonnet 5.5high35.5%65.62026-09-28
17Gemini 3.8 Flashhigh35.0%64.42026-09-02
18Claude Opus 5high30.2%67.92026-07-24
19Gemini 3.8 Flashmedium24.2%64.42026-09-02
20GPT-6 Solmax23.0%66.42026-09-22
21Gemini 3.8 Flashlow15.1%61.02026-09-02
22Grok 4.7high10.0%63.92026-09-21
23GPT-6 Solxhigh9.5%64.32026-09-22
24Grok 4.7medium8.4%–2026-09-21
25GPT-5.6 Solmax7.8%68.02026-07-09
26GPT-5.6 Solxhigh7.0%66.62026-07-09
27Grok 4.7xhigh7.0%61.62026-09-21
28GPT-6 Solhigh5.9%63.72026-09-22
29GPT-6 Solmedium4.9%62.62026-09-22
30Claude Sonnet 5.5medium4.9%58.22026-09-28
31DeepSeek V4.1 Flashlow3.3%–2026-09-10
32DeepSeek V4.1 Flashhigh3.2%55.62026-09-10
33Claude Haiku 5.5max2.8%60.52026-10-07
34DeepSeek V4.1 Flashmax2.7%61.22026-09-10
35GPT-5.6 Solhigh2.2%66.32026-07-09
36Grok 4.6xhigh2.1%62.52026-08-11
37GPT-6 Sollow1.8%59.72026-09-22
38Claude Opus 4.8high1.5%60.72026-06-01
39GPT-6 Solno reasoning1.2%52.82026-09-22
40GPT-5.6 Solmedium1.1%64.42026-07-09
41GPT-5.6 Terramax0.8%64.42026-07-09
42Grok 4.7low0.8%61.02026-09-21
43GPT-5.6 Terraxhigh0.7%62.82026-07-09
44GPT-6 Lunamax0.6%60.02026-09-22
45Claude Sonnet 5.5low0.6%58.02026-09-28
46GPT-5.6 Terrahigh0.5%61.32026-07-09
47GPT-6 Lunaxhigh0.5%58.72026-09-22
48Claude Opus 4.6max0.5%57.42025-12-17
49DeepSeek V4.1 Flashno reasoning0.5%48.02026-09-10
50GPT-5.5high0.4%65.32026-04-23
51Gemini 3.1 Pro0.4%60.62026-03-05
52GPT-6 Lunahigh0.4%57.32026-09-22
53GPT-5.6 Sollow0.3%62.22026-07-09
54Grok 4.5high0.3%59.32026-07-16
55GPT-6 Lunamedium0.3%56.32026-09-22
56Grok 4.5low0.3%–2026-07-16
57Grok 4.5medium0.3%–2026-07-16
58Claude Opus 4.7high0.2%61.02026-04-16
59GPT-5.6 Lunamax0.2%59.92026-07-09
60GPT-5.4high0.2%57.92026-03-05

More reasoning benchmarks · All LLM benchmarks

Cite as: BenchLeader, “ARC-AGI-3 leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/arc_agi_3.

ARC-AGI-3: questions

What does ARC-AGI-3 measure?
ARC-AGI-3 moves to interactive games: the agent must explore an environment, discover its rules and reach goals without instructions, measuring learning efficiency over many steps. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads ARC-AGI-3?
GPT-6 Astra leads ARC-AGI-3 with 99.9% as of 24 Oct 2026, ahead of GPT-6 Astra at 98.6%.
How many models have ARC-AGI-3 results?
69 model configurations have a ARC-AGI-3 result on BenchLeader, all taken from ARC Prize.
Who runs ARC-AGI-3 and how often is it updated?
ARC-AGI-3 is published by ARC Prize. BenchLeader re-reads the published results once a day and records the date each result was published.
Does ARC-AGI-3 count toward the BenchLeader Index?
Yes. ARC-AGI-3 contributes to the reasoning category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.