BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

ARC-AGI-2

Semi-private ARC-AGI-2 evaluation set.

As of 24 Oct 2026, GPT-6 Astra leads ARC-AGI-2 on BenchLeader with 95.0%, ahead of GPT-6.1 Sol at 94.2%, across 256 model configurations with a published result.

Published by
ARC Prize
Category
Reasoning
Index weight
1.0
Models
256
Data as of
24 Oct 2026
Newest result
17 Oct 2026

ARC Prize Foundation, arcprize.org.

What the test looks like

ARC-AGI-2 keeps the format of ARC-AGI-1 but with tasks designed to resist brute-force search and pattern matching: symbolic interpretation, compositional rules and contextual rule application. Humans solve every task; models started near zero.

How it is scored

Percent of semi-private tasks solved with two attempts, run by the ARC Prize Foundation with cost per task.

What to keep in mind

Highly sensitive to reasoning budget and to the harness; compare configurations at similar cost. Scores change quickly as labs target it.

Top 15
  1. 1GPT-6 Astra (max)95.0%
  2. 2GPT-6.1 Sol (max)94.2%
  3. 3Claude Opus 5.5 (high)93.3%
  4. 4GPT-6 Astra (xhigh)93.3%
  5. 5Claude Opus 5.5 (xhigh)92.5%
  6. 6GPT-5.6 Sol (max)92.5%
  7. 7GPT-6 Astra (high)92.1%
  8. 8GPT-6 Astra (medium)92.1%
  9. 9Claude Opus 5.5 (max)91.7%
  10. 10GPT-6.1 Sol (high)91.7%
  11. 11GPT-6.1 Sol (xhigh)91.7%
  12. 12Gemini 4 Argon (high)91.3%
  13. 13Claude Opus 5 (max)90.4%
  14. 14Claude Fable 5.1 (max)90.0%
  15. 15Claude Fable 5.1 (xhigh)90.0%
256 of 256
#
1GPT-6 Astramax95.0%70.42026-09-02
2GPT-6.1 Solmax94.2%69.92026-09-29
3Claude Opus 5.5high93.3%69.42026-09-22
4GPT-6 Astraxhigh93.3%68.92026-09-02
5Claude Opus 5.5xhigh92.5%70.02026-09-22
6GPT-5.6 Solmax92.5%68.02026-07-09
7GPT-6 Astrahigh92.1%69.62026-09-02
8GPT-6 Astramedium92.1%67.52026-09-02
9Claude Opus 5.5max91.7%72.62026-09-22
10GPT-6.1 Solhigh91.7%67.72026-09-29
11GPT-6.1 Solxhigh91.7%67.72026-09-29
12Gemini 4 Argonhigh91.3%70.82026-10-13
13Claude Opus 5max90.4%68.82026-07-24
14Claude Fable 5.1max90.0%70.52026-09-01
15Claude Fable 5.1xhigh90.0%69.52026-09-01
16GPT-5.6 Solxhigh90.0%66.62026-07-09
17GPT-6 Solmax89.6%66.42026-09-22
18Claude Fable 5max89.2%69.32026-06-09
19Gemini 3.8 Flashhigh89.2%64.42026-09-02
20Claude Fable 5.1high88.8%69.42026-09-01
21Claude Opus 5high88.3%67.92026-07-24
22Claude Fable 5xhigh88.3%–2026-06-09
23Claude Opus 5.5medium87.5%66.52026-09-22
24Claude Fable 5high87.5%64.62026-06-09
25Gemini 4 Argonmedium87.1%–2026-10-13
26GPT-6.1 Solmedium86.7%67.22026-09-29
27Claude Fable 5.1medium86.3%67.32026-09-01
28GPT-5.6 Solhigh85.4%66.32026-07-09
29GPT-6 Astralow85.4%65.92026-09-02
30GPT-5.5xhigh85.0%66.12026-04-22
31Gemini 3.7 Flashhigh84.6%63.52026-08-13
32Gemini 3 Deep Thinkthinking84.6%–2026-02-12
33GPT-5.5 Prohigh84.6%–2026-04-23
34GPT-5.5 Proxhigh84.2%65.32026-04-23
35Claude Sonnet 5.5xhigh83.9%67.92026-09-28
36GPT-5.6 Terramax83.9%64.42026-07-09
37GPT-5.5high83.3%65.32026-04-22
38GPT-5.4 Proxhigh83.3%61.92026-03-04
39Gemini 3.8 Flashmedium82.9%64.42026-09-02
40Claude Fable 5medium82.5%–2026-06-09
41Claude Fable 5.1low78.3%65.02026-09-01
42GPT-6 Solxhigh78.1%64.32026-09-22
43NewDeepSeek V4.1 Flash 1017max78.1%64.02026-10-17
44Gemini 3.8 Flashlow77.5%61.02026-09-02
45Gemini 3.1 Pro77.1%60.62026-02-19
46Claude Fable 5low76.8%–2026-06-09
47Dots3 Notemax76.8%–2026-08-14
48GPT-6.1 Sollow76.7%64.92026-09-29
49Claude 4.7max75.8%–2026-04-16
50GPT-5.6 Terraxhigh74.2%62.82026-07-09
51GPT-5.4xhigh74.0%64.72026-03-04
52NewDeepSeek V4.1 Flash 1017low73.6%–2026-10-17
53DeepSeek V4.1 Flashmax72.9%61.22026-09-10
54GPT-5.272.9%58.02026-02-03
55Gemini 3.5 Flashhigh72.1%63.12026-05-19
56Claude Opus 4.8high72.1%60.72026-06-01
57Claude Opus 4.8medium71.7%–2026-06-01
58GPT-5.5medium70.4%62.92026-04-22
59Claude Opus 5.5low70.1%64.72026-09-22
60Claude Opus 4.6high69.2%62.72026-02-05

More reasoning benchmarks · All LLM benchmarks

Cite as: BenchLeader, “ARC-AGI-2 leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/arc_agi_2.

ARC-AGI-2: questions

What does ARC-AGI-2 measure?
ARC-AGI-2 keeps the format of ARC-AGI-1 but with tasks designed to resist brute-force search and pattern matching: symbolic interpretation, compositional rules and contextual rule application. Humans solve every task; models started near zero. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads ARC-AGI-2?
GPT-6 Astra leads ARC-AGI-2 with 95.0% as of 24 Oct 2026, ahead of GPT-6.1 Sol at 94.2%.
How many models have ARC-AGI-2 results?
256 model configurations have a ARC-AGI-2 result on BenchLeader, all taken from ARC Prize.
Who runs ARC-AGI-2 and how often is it updated?
ARC-AGI-2 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-2 count toward the BenchLeader Index?
Yes. ARC-AGI-2 contributes to the reasoning category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.