Vision & documents

AI2D_TEST leaderboard

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

As of 2026-10-10, the highest measured score on AI2D_TEST is 94.1% by Gemini 3 Pro. 59 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	GPT-5.6 Sol	94.7%	estimated ± 2.0 pp, low confidence
2	Gemini 3 Pro	94.1%	measured
3	Qwen3.5 397B	93.9%	measured
4	GPT-5.5	93.9%	estimated ± 2.0 pp, low confidence
5	dots3-note Preview	93.7%	estimated ± 1.0 pp, medium confidence
6	GPT-5.6 Terra	93.6%	estimated ± 2.0 pp, medium confidence
7	Claude Mythos 5	93.5%	estimated ± 0.9 pp, low confidence
8	Qwen3.8 Max	93.5%	estimated ± 0.9 pp, low confidence
9	Qwen3.8-Omni-Flash	93.5%	estimated ± 0.9 pp, low confidence
10	Kimi K3	93.5%	estimated ± 0.9 pp, low confidence
11	Claude Opus 4.7 (Adaptive)	93.5%	estimated ± 0.9 pp, low confidence
12	Qwen3.8-Flash-Next	93.5%	estimated ± 0.9 pp, low confidence
13	Qwen3.8-27B	93.5%	estimated ± 0.9 pp, low confidence
14	Claude Opus 4.8	93.5%	estimated ± 0.9 pp, low confidence
15	GLM-5.3-Flash	93.5%	estimated ± 0.9 pp, low confidence
16	Gemini 3.7 Flash	93.5%	estimated ± 0.9 pp, low confidence
17	Muse Spark 1.1	93.5%	estimated ± 0.9 pp, low confidence
18	Claude Sonnet 5	93.5%	estimated ± 0.9 pp, low confidence
19	Sakana Fugu-Ultra	93.5%	estimated ± 0.9 pp, low confidence
20	Muse Spark	93.5%	estimated ± 0.9 pp, low confidence
21	Qwen3.7 Plus	93.5%	estimated ± 0.9 pp, low confidence
22	Seed 2.1 Pro	93.5%	estimated ± 0.9 pp, low confidence
23	Sakana Fugu	93.5%	estimated ± 0.9 pp, low confidence
24	Gemini 3.5 Flash	93.5%	estimated ± 0.9 pp, low confidence
25	GPT-5.4	93.5%	estimated ± 0.9 pp, low confidence
26	Seed 2.1 Turbo	93.5%	estimated ± 0.9 pp, low confidence
27	Inkling	93.5%	estimated ± 0.9 pp, medium confidence
28	Qwen3.6 Plus	93.4%	estimated ± 0.9 pp, medium confidence
29	Inkling-Small	93.4%	estimated ± 0.9 pp, medium confidence
30	MiMo-V2.5	93.4%	estimated ± 0.9 pp, medium confidence
31	Kimi K2.6	93.4%	estimated ± 0.9 pp, medium confidence
32	Gemini 3.1 Pro	93.4%	estimated ± 0.9 pp, medium confidence
33	Qwen3.5-27B	93.4%	estimated ± 0.9 pp, medium confidence
34	Muse Glimmer 30B	93.2%	estimated ± 0.9 pp, medium confidence
35	Qwen3.6-27B	92.9%	estimated ± 0.9 pp, medium confidence
36	Qwen3.6-35B-A3B	92.7%	measured
37	GPT-5.6 Luna	92.5%	estimated ± 2.0 pp, medium confidence
38	Grok 4.3	92.3%	estimated ± 2.0 pp, medium confidence
39	Pareto 26.9	92.3%	estimated ± 2.0 pp, medium confidence
40	GPT-5.2	92.2%	measured
41	Claude Opus 4.6	91.9%	estimated ± 2.0 pp, medium confidence
42	MiniMax M3	91.8%	estimated ± 1.3 pp, medium confidence
43	Gemma 4 31B	91.7%	estimated ± 2.0 pp, medium confidence
44	GPT-5.4 mini	91.6%	estimated ± 2.0 pp, medium confidence
45	Step 5 Preview	91.3%	estimated ± 2.0 pp, medium confidence
46	Kimi K2.5 (Reasoning)	91.1%	estimated ± 0.9 pp, medium confidence
47	Qwen3.5-35B-A3B	91.1%	estimated ± 0.9 pp, medium confidence
48	Claude Sonnet 4.6	90.8%	estimated ± 0.9 pp, medium confidence
49	Kimi K2.5	90.8%	measured
50	Qwen3.5-122B-A10B	90.3%	estimated ± 0.9 pp, medium confidence
51	Gemma 4 26B A4B	90.1%	estimated ± 2.0 pp, medium confidence
52	Gemma 4 12B	90.1%	estimated ± 1.0 pp, medium confidence
53	Ternary Bonsai 2 27B	88.7%	estimated ± 1.4 pp, medium confidence
54	Interfaze Beta	88.6%	estimated ± 2.0 pp, medium confidence
55	Nemotron 3 Nano Omni 30B A3B	88.5%	measured
56	Gemini 3.1 Flash-Lite	87.8%	estimated ± 0.9 pp, medium confidence
57	Grok 4.20	87.8%	estimated ± 0.9 pp, low confidence
58	Command A+	87.8%	estimated ± 0.9 pp, low confidence
59	Claude Opus 4.5	87.7%	measured
60	LFM2.5-VL-3B	87.6%	estimated ± 1.4 pp, medium confidence
61	LFM2.5-VL-450M	87.6%	estimated ± 1.4 pp, low confidence
62	North Micro Vision Instruct	87.6%	estimated ± 1.4 pp, low confidence
63	ZAYA1-VL-8B	87.5%	measured
64	GPT-5.4 nano	85.7%	estimated ± 2.0 pp, low confidence
65	Gemma 4 E4B	76.6%	estimated ± 2.0 pp, low confidence
66	Qwen2.5-VL-32B	74.1%	estimated ± 2.0 pp, low confidence
67	Gemma 4 E2B	69.6%	estimated ± 2.0 pp, low confidence

AI2D_TEST leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General