BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases9

ARC-AGI-1

Semi-private ARC-AGI-1 evaluation set.

As of 25 Oct 2026, Gemini 4 Argon leads ARC-AGI-1 on BenchLeader with 98.5%, ahead of GPT-6 Astra at 98.5%, across 256 model configurations with a published result.

Published by
ARC Prize
Category
Reasoning
Index weight
0.5
Models
256
Data as of
25 Oct 2026
Newest result
17 Oct 2026

ARC Prize Foundation, arcprize.org.

What the test looks like

ARC-AGI-1 shows a few examples of a grid transformation and asks for the output on a new grid, testing the ability to infer a rule from very few examples. The semi-private evaluation set is withheld from the public.

How it is scored

Percent of tasks solved with at most two attempts, run by the ARC Prize Foundation; the cost per task is reported alongside.

What to keep in mind

Scores rise sharply with test-time compute, so the same model at different reasoning budgets can differ by tens of points; efficiency matters as much as the score. Largely solved by top systems now.

Top 15
  1. 1Gemini 4 Argon (high)98.5%
  2. 2GPT-6 Astra (high)98.5%
  3. 3Claude Opus 5.5 (high)98.5%
  4. 4Claude Fable 5 (max)98.5%
  5. 5GPT-6 Astra (xhigh)98.5%
  6. 6GPT-6.1 Sol (high)98.5%
  7. 7GPT-6.1 Sol (xhigh)98.5%
  8. 8Gemini 3.8 Flash (high)98.5%
  9. 9Gemini 4 Argon (medium)98.5%
  10. 10Claude Fable 5 (xhigh)98.5%
  11. 11Gemini 3.1 Pro98.0%
  12. 12Claude Opus 5.5 (max)97.5%
  13. 13Claude Fable 5.1 (max)97.5%
  14. 14GPT-6 Astra (max)97.5%
  15. 15Claude Opus 5.5 (xhigh)97.5%
256 of 256
#
1Gemini 4 Argonhigh98.5%70.72026-10-13
2GPT-6 Astrahigh98.5%69.62026-09-02
3Claude Opus 5.5high98.5%69.42026-09-22
4Claude Fable 5max98.5%69.32026-06-09
5GPT-6 Astraxhigh98.5%68.92026-09-02
6GPT-6.1 Solhigh98.5%67.72026-09-29
7GPT-6.1 Solxhigh98.5%67.72026-09-29
8Gemini 3.8 Flashhigh98.5%64.42026-09-02
9Gemini 4 Argonmedium98.5%–2026-10-13
10Claude Fable 5xhigh98.5%–2026-06-09
11Gemini 3.1 Pro98.0%60.52026-02-19
12Claude Opus 5.5max97.5%72.62026-09-22
13Claude Fable 5.1max97.5%70.52026-09-01
14GPT-6 Astramax97.5%70.42026-09-02
15Claude Opus 5.5xhigh97.5%70.02026-09-22
16Claude Opus 5max97.5%68.82026-07-24
17Claude Opus 5high97.5%67.92026-07-24
18GPT-6 Astramedium97.5%67.52026-09-02
19GPT-5.6 Solxhigh97.5%66.62026-07-09
20Claude Opus 5.5medium97.5%66.52026-09-22
21Gemini 3.8 Flashmedium97.5%64.42026-09-02
22GPT-5.6 Solhigh97.0%66.32026-07-09
23GPT-6.1 Solmax96.5%69.92026-09-29
24Claude Fable 5.1xhigh96.5%69.52026-09-01
25Claude Sonnet 5.5max96.5%69.42026-09-28
26GPT-5.6 Solmax96.5%68.02026-07-09
27Claude Sonnet 5.5xhigh96.5%67.92026-09-28
28GPT-6 Astralow96.5%65.92026-09-02
29GPT-5.6 Terramax96.5%64.42026-07-09
30GPT-5.5 Prohigh96.5%–2026-04-23
31Claude Fable 5.1high96.0%69.42026-09-01
32Gemini 3 Deep Thinkthinking96.0%–2026-02-12
33GPT-6.1 Solmedium95.5%67.22026-09-29
34GPT-6 Solmax95.5%66.42026-09-22
35Claude Fable 5high95.5%64.42026-06-09
36Gemini 3.7 Flashhigh95.5%63.52026-08-13
37GPT-5.5xhigh95.0%66.12026-04-22
38GPT-5.5 Proxhigh95.0%65.32026-04-23
39Claude Fable 5.1medium94.5%67.32026-09-01
40GPT-5.5high94.5%65.32026-04-22
41Kimi K3max94.5%64.82026-07-16
42GPT-5.4 Proxhigh94.5%61.92026-03-04
43DeepSeek V4.1 Flashmax94.5%61.22026-09-10
44GPT-5.294.5%58.02026-02-03
45DeepSeek V4.1 Flash 1017high94.5%–2026-10-17
46GPT-5.6 Terraxhigh94.0%62.82026-07-09
47Claude Opus 4.6high94.0%62.72026-02-05
48GPT-5.4xhigh93.7%64.72026-03-04
49GPT-6.1 Sollow93.5%64.92026-09-29
50DeepSeek V4.1 Flash 1017max93.5%64.02026-10-17
51Claude 4.7high93.5%–2026-04-16
52Claude Opus 4.6max93.0%57.42026-02-05
53GPT-6 Solxhigh92.7%64.32026-09-22
54GPT-5.4high92.7%58.92026-03-04
55GPT-5.6 Solmedium92.5%64.42026-07-09
56Claude Opus 4.8max92.5%63.32026-06-01
57Gemini 3.5 Flashhigh92.5%63.12026-05-19
58Claude Fable 5medium92.5%–2026-06-09
59GPT-5.5medium92.2%62.92026-04-22
60Claude Sonnet 5.5high92.0%65.62026-09-28

More reasoning benchmarks · All LLM benchmarks

Cite as: BenchLeader, “ARC-AGI-1 leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/arc_agi_1.

ARC-AGI-1: questions

What does ARC-AGI-1 measure?
ARC-AGI-1 shows a few examples of a grid transformation and asks for the output on a new grid, testing the ability to infer a rule from very few examples. The semi-private evaluation set is withheld from the public. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads ARC-AGI-1?
Gemini 4 Argon leads ARC-AGI-1 with 98.5% as of 25 Oct 2026, ahead of GPT-6 Astra at 98.5%.
How many models have ARC-AGI-1 results?
256 model configurations have a ARC-AGI-1 result on BenchLeader, all taken from ARC Prize.
Who runs ARC-AGI-1 and how often is it updated?
ARC-AGI-1 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-1 count toward the BenchLeader Index?
Yes. ARC-AGI-1 contributes to the reasoning category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.