Skip to content

Benchmark

GAIA

Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.

Models measured
71
Scores recorded
291
Sources
3
IQ weight
1.01

Score distribution

37%
Mean
32.2%
Median
28.6%
Std dev
2.3%
10th percentile
80.6%
90th percentile
0%
Min
0%94.6%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.01 for chat models, measured over 71 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Claude Opus 5.5Level 1· GAIA Leaderboard data94.6%
2Claude Opus 4.8Level 1· GAIA Leaderboard data93.5%
3DSV4Level 1· GAIA Leaderboard data93.5%
4GPT 5.5Level 1· GAIA Leaderboard data93.5%
5GPT-6.1 SolLevel 1· GAIA Leaderboard data91.4%
6Kimi K3Level 1· GAIA Leaderboard data91.4%
7GLM 5.2Level 1· GAIA Leaderboard data89.2%
8Claude Opus 4.8· GAIA Leaderboard data88.7%
9Claude Opus 5.5Level 2· GAIA Leaderboard data88.7%
10Claude Opus 5.5· GAIA Leaderboard data88.4%
11DSV4· GAIA Leaderboard data88%
12GPT 5.4Level 1· GAIA Leaderboard data87.1%
13Claude Opus 4.8Level 2· GAIA Leaderboard data86.8%
14Gemini 2.5 ProLevel 1· GAIA Leaderboard data86.8%
15Manus· Steel.dev Agent Leaderboards data86.5%
16GPT 5.5· GAIA Leaderboard data86.4%
17DSV4Level 2· GAIA Leaderboard data86.2%
18Gemini 3 Pro PreviewLevel 1· GAIA Leaderboard data86%
19Qwen3.8 FlashLevel 1· GAIA Leaderboard data86%
20Claude Opus 4.8Level 3· GAIA Leaderboard data85.7%
21GPT 5.5Level 2· GAIA Leaderboard data85.5%
22GPT 5.6Level 1· GAIA Leaderboard data84.9%
23Step 3.5 FlashStep 3.5 Flash official headline table · vanilla · no file· Reported by StepFun· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.84.5%
24GPT-6.1 Sol· GAIA Leaderboard data84.4%
25GPT-6.1 SolLevel 2· GAIA Leaderboard data84.3%
26Claude Opus 4.5Level 1· GAIA Leaderboard data83.9%
27DeepSeek V4 ProLevel 1· GAIA Leaderboard data83.9%
28DSV4Level 3· GAIA Leaderboard data83.7%
29GPT-5.4 miniLevel 1· GAIA Leaderboard data81.7%
30GPT-5.6 TerraLevel 1· GAIA Leaderboard data80.6%
31Kimi K3· GAIA Leaderboard data80.4%
32GLM 5.2· GAIA Leaderboard data79.7%
33Kimi K3Level 2· GAIA Leaderboard data79.2%
34GLM 5.2Level 2· GAIA Leaderboard data78.6%
35DeepSeek FlashLevel 1· GAIA Leaderboard data78.5%
36Gemini 3 Pro Preview· GAIA Leaderboard data78.1%
37Gemini 2.5 Pro· GAIA Leaderboard data77%
38Gemini 2.5 ProLevel 2· GAIA Leaderboard data76.7%
39Claude Opus 5.5Level 3· GAIA Leaderboard data75.5%
40GPT 5.5Level 3· GAIA Leaderboard data75.5%
41Gemini 3 Pro PreviewLevel 2· GAIA Leaderboard data75.5%
42DeepSeek V4 FlashLevel 1· GAIA Leaderboard data75.3%
43DeepSeek V4 Flash 0731Level 1· GAIA Leaderboard data75.3%
44Claude Opus 4.5Level 2· GAIA Leaderboard data74.2%
45GPT-5.6 LunaLevel 1· GAIA Leaderboard data74.2%
46Claude Opus 4.5· GAIA Leaderboard data74.1%
47Qwen3.8 Flash· GAIA Leaderboard data73.8%
48Qwen3.8 FlashLevel 2· GAIA Leaderboard data73.6%
49GPT 5.6· GAIA Leaderboard data73.4%
50Claude Sonnet 4.6Level 1· GAIA Leaderboard data72%

Showing 1–50 of 291 results.

Data credits: Steel.dev Agent Leaderboards, MIT (licence) · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links.