Vision & documents

CountBench leaderboard

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

As of 2026-10-10, the highest measured score on CountBench is 97.8% by Qwen3.6-27B. 59 more models have estimated scores, calibrated from the benchmarks they were measured on.

Measured scores: benchlm.ai.

#	Model	Score	Source
1	Claude Mythos 5	99.4%	estimated ± 3.7 pp, low confidence
2	Claude Opus 4.7 (Adaptive)	98.8%	estimated ± 3.7 pp, low confidence
3	Claude Opus 4.8	98.5%	estimated ± 3.7 pp, low confidence
4	GLM-5.3-Flash	98.3%	estimated ± 3.7 pp, low confidence
5	Gemini 3.7 Flash	98.1%	estimated ± 3.7 pp, low confidence
6	Muse Spark 1.1	98.1%	estimated ± 3.7 pp, low confidence
7	Claude Sonnet 5	98.0%	estimated ± 3.7 pp, low confidence
8	Qwen3.6-27B	97.8%	measured
9	Sakana Fugu-Ultra	97.5%	estimated ± 3.7 pp, low confidence
10	Step 3.7 Flash	97.5%	estimated ± 0.9 pp, medium confidence
11	Gemini 3.1 Pro	97.5%	estimated ± 2.2 pp, medium confidence
12	Qwen3.8 Max	97.4%	estimated ± 0.8 pp, low confidence
13	Kimi K3	97.4%	estimated ± 0.8 pp, low confidence
14	Seed 2.1 Pro	97.4%	estimated ± 0.8 pp, low confidence
15	Qwen3.8-Omni-Flash	97.4%	estimated ± 0.8 pp, low confidence
16	Qwen3.8-Flash-Next	97.4%	estimated ± 0.8 pp, low confidence
17	Qwen3.7 Plus	97.4%	estimated ± 0.8 pp, low confidence
18	Seed 2.1 Turbo	97.4%	estimated ± 0.8 pp, low confidence
19	Qwen3.8-27B	97.4%	estimated ± 0.8 pp, low confidence
20	Qwen3.6 Plus	97.3%	estimated ± 0.8 pp, medium confidence
21	dots3-note Preview	97.3%	estimated ± 0.8 pp, medium confidence
22	Gemini 3 Pro	97.3%	measured
23	Kimi K2.6	97.3%	estimated ± 0.8 pp, medium confidence
24	Qwen3.6-35B-A3B	97.3%	estimated ± 1.6 pp, medium confidence
25	Qwen3.5 397B	97.2%	measured
26	Sakana Fugu	97.1%	estimated ± 3.7 pp, low confidence
27	Qwen3.5-122B-A10B	97.0%	estimated ± 0.8 pp, medium confidence
28	Qwen3.5-27B	96.9%	estimated ± 0.8 pp, medium confidence
29	Gemini 3.5 Flash	96.8%	estimated ± 3.7 pp, low confidence
30	MiMo-V2.5	96.7%	estimated ± 1.8 pp, medium confidence
31	GPT-5.4	96.2%	estimated ± 2.2 pp, medium confidence
32	Inkling	96.2%	estimated ± 3.7 pp, low confidence
33	Muse Spark	96.0%	estimated ± 2.2 pp, medium confidence
34	Inkling-Small	95.9%	estimated ± 3.7 pp, low confidence
35	GPT-5.6 Sol	95.8%	estimated ± 5.5 pp, low confidence
36	GPT-5.5	95.5%	estimated ± 5.5 pp, low confidence
37	GPT-5.6 Terra	95.4%	estimated ± 5.5 pp, low confidence
38	GPT-5.6 Luna	95.0%	estimated ± 5.5 pp, low confidence
39	Grok 4.3	95.0%	estimated ± 5.5 pp, low confidence
40	Pareto 26.9	95.0%	estimated ± 5.5 pp, low confidence
41	Gemma 4 31B	94.8%	estimated ± 5.5 pp, low confidence
42	GPT-5.4 mini	94.7%	estimated ± 5.5 pp, low confidence
43	Kimi K2.5 (Reasoning)	94.7%	estimated ± 3.7 pp, low confidence
44	Claude Sonnet 4.6	94.6%	estimated ± 3.7 pp, low confidence
45	Step 5 Preview	94.6%	estimated ± 5.5 pp, low confidence
46	Gemma 4 26B A4B	94.3%	estimated ± 5.5 pp, low confidence
47	Kimi K2.5	94.1%	measured
48	Interfaze Beta	93.8%	estimated ± 5.5 pp, low confidence
49	Qwen3.5-35B-A3B	93.5%	estimated ± 0.8 pp, medium confidence
50	Ternary Bonsai 2 27B	93.4%	estimated ± 3.8 pp, high confidence
51	Gemini 3.1 Flash-Lite	93.2%	estimated ± 3.7 pp, low confidence
52	GPT-5.4 nano	93.0%	estimated ± 5.5 pp, low confidence
53	Grok 4.20	92.6%	estimated ± 2.2 pp, medium confidence
54	GPT-5.2	91.9%	measured
55	Claude Opus 4.6	91.9%	estimated ± 2.2 pp, medium confidence
56	MiniMax M3	91.0%	estimated ± 1.8 pp, medium confidence
57	Gemma 4 E4B	90.7%	estimated ± 5.5 pp, low confidence
58	Gemma 4 12B	90.6%	estimated ± 0.8 pp, medium confidence
59	Claude Opus 4.5	90.6%	measured
60	Nemotron 3 Nano Omni 30B A3B	90.3%	estimated ± 2.0 pp, medium confidence
61	Qwen2.5-VL-32B	90.1%	estimated ± 5.5 pp, low confidence
62	Gemma 4 E2B	89.2%	estimated ± 5.5 pp, low confidence
63	ZAYA1-VL-8B	88.1%	measured
64	LFM2.5-VL-3B	87.3%	measured
65	Command A+	83.7%	estimated ± 3.7 pp, low confidence
66	North Micro Vision Instruct	81.0%	estimated ± 3.8 pp, high confidence
67	Muse Glimmer 30B	77.2%	estimated ± 2.5 pp, low confidence
68	LFM2.5-VL-450M	73.3%	measured

CountBench leaderboard

All leaderboards

Agentic · terminal

Agentic · tools

Coding

Math

Knowledge & reasoning

Instruction following

Multilingual

Vision & documents

Long context

General