上下文窗口效率探索器

分析300个AI模型的评分/上下文令牌比率，找出最充分利用上下文窗口、输出容量和成本的模型。

上下文窗口 vs 评分

LMMarketCap.com

效率概览

所有分析模型的关键效率指标。

最高效 (128K+)

GPT-5.2 Chat

707.0 score/MToken

最佳输出效率

Gemma 2 27B

37.8 score/1K output

最佳成本效率

gpt-oss-20b

1063.0 score/$

平均总体效率

6.0%

所有模型标准化

效率排名

按每百万上下文令牌评分排名的前50个模型。

#	模型	提供商	评分	上下文	输出	评分/百万令牌	输出效率	成本效率	层级
1	Gemma 2 27BGoogle	Google	77	8K	2K	9448.2	37.8	119.1	Small
2	GPT-4OpenAI	OpenAI	65	8K	4K	7923.3	15.8	1.4	Small
3	R1 Distill Llama 70BDeepSeek	DeepSeek	41	8K	7K	4968.3	5.5	50.9	Small
4	Hy-MT2-1.8BTencent	Tencent	40	8K	4K	4882.8	9.8	362.0	Small
5	Hy-MT2-30B-A3BTencent	Tencent	40	8K	4K	4882.8	9.8	216.8	Small
6	Hy-MT2-7BTencent	Tencent	40	8K	4K	4882.8	9.8	216.8	Small
7	Phi 4Microsoft	Microsoft	60	16K	15K	3674.3	4.1	573.3	Small
8	GLM 5.2 (free)Zhipu AI	Zhipu AI	76	33K	29K	2310.2	2.6	Free	Medium
9	R1DeepSeek	DeepSeek	74	64K	16K	1153.1	4.6	46.1	Medium
10	MiniMax M2-herMiniMax	MiniMax	68	66K	2K	1039.1	33.3	90.8	Medium
11	Mixtral 8x22B InstructMistral AI	Mistral AI	63	66K	52K	967.4	1.2	15.8	Medium
12	GLM 4.5VZhipu AI	Zhipu AI	62	66K	16K	950.6	3.8	51.9	Medium
13	Qwen3 30B A3B Thinking 2507Alibaba	Alibaba	64	82K	33K	782.5	2.0	49.3	Medium
14	GPT-5.2 ChatOpenAI	OpenAI	91	128K	32K	707.0	2.8	11.5	Large
15	LFM2.5-2.6B (free)Liquid AI	Liquid AI	40	66K	8K	610.4	4.9	Free	Medium
16	GLM 4.5Zhipu AI	Zhipu AI	75	131K	98K	573.0	0.8	53.6	Large
17	GPT-4o (batch)OpenAI	OpenAI	72	128K	16K	562.5	4.4	23.0	Large
18	GPT-4o (2024-11-20)OpenAI	OpenAI	71	128K	16K	556.3	4.3	11.4	Large
19	GPT-4o (2024-08-06)OpenAI	OpenAI	71	128K	16K	556.3	4.3	11.4	Large
20	GPT-4oOpenAI	OpenAI	71	128K	16K	556.3	4.3	11.4	Large
21	GPT-4o (2024-05-13)OpenAI	OpenAI	71	128K	4K	556.3	17.4	7.1	Large
22	GPT-4o-miniOpenAI	OpenAI	69	128K	16K	541.4	4.2	184.8	Large
23	GLM 4.5 AirZhipu AI	Zhipu AI	71	131K	98K	539.4	0.7	144.3	Large
24	Qwen3 VL 235B A22B ThinkingAlibaba	Alibaba	69	131K	33K	528.7	2.1	31.5	Large
25	GPT-4 TurboOpenAI	OpenAI	67	128K	4K	521.1	16.3	3.3	Large
26	Mistral LargeMistral AI	Mistral AI	66	128K	102K	514.8	0.6	16.5	Large
27	Llama 3.3 70B InstructMeta	Meta	67	131K	16K	509.6	4.1	318.1	Large
28	DeepSeek V3.2DeepSeek	DeepSeek	83	164K	66K	509.0	1.3	249.3	Large
29	Mistral Large 2407Mistral AI	Mistral AI	66	131K	105K	502.8	0.6	16.5	Large
30	GLM 4.6VZhipu AI	Zhipu AI	66	131K	33K	500.5	2.0	109.3	Large
31	Qwen3 235B A22B Thinking 2507Alibaba	Alibaba	65	131K	118K	498.2	0.6	51.6	Large
32	Llama 3.1 70B InstructMeta	Meta	65	131K	16K	498.2	4.0	163.3	Large
33	Qwen3 30B A3BAlibaba	Alibaba	64	131K	16K	489.0	3.9	206.8	Large
34	GPT-4o-mini (batch)OpenAI	OpenAI	63	128K	16K	488.3	3.8	333.3	Large
35	R1 0528DeepSeek	DeepSeek	79	164K	33K	484.6	2.4	59.9	Large
36	GPT-4 Turbo (batch)OpenAI	OpenAI	62	128K	4K	480.5	15.0	6.2	Large
37	Mercury 2Inception	Inception	61	128K	50K	475.8	1.2	121.8	Large
38	gpt-oss-120bOpenAI	OpenAI	62	131K	66K	470.0	0.9	164.3	Large
39	Qwen3 8BAlibaba	Alibaba	61	131K	8K	465.4	7.4	213.3	Large
40	GPT-4o-mini (2024-07-18)OpenAI	OpenAI	57	128K	16K	441.4	3.4	150.7	Large
41	DeepSeek V3.2 ExpDeepSeek	DeepSeek	72	164K	66K	438.2	1.1	211.2	Large
42	DeepSeek V3.1DeepSeek	DeepSeek	72	164K	33K	438.2	2.2	119.7	Large
43	DeepSeek V3 0324DeepSeek	DeepSeek	72	164K	147K	438.2	0.5	114.9	Large
44	gpt-oss-20bOpenAI	OpenAI	57	131K	33K	437.9	1.8	1063.0	Large
45	gpt-oss-20b (batch)OpenAI	OpenAI	57	131K	118K	437.9	0.5	844.1	Large
46	o3 ProOpenAI	OpenAI	87	200K	100K	433.5	0.9	1.7	Large
47	o3OpenAI	OpenAI	87	200K	100K	433.5	0.9	17.3	Large
48	o3 (batch)OpenAI	OpenAI	87	200K	100K	433.5	0.9	34.7	Large
49	Claude Opus 4.5Anthropic	Anthropic	85	200K	64K	425.5	1.3	5.7	Large
50	Claude Opus 4.5 (batch)Anthropic	Anthropic	85	200K	64K	425.5	1.3	11.3	Large

层级分析

不同上下文窗口层级的效率分析。

Small7 models

平均评分52

评分/百万令牌5808.9

最佳

Gemma 2 27B

最差

Phi 4

Medium7 models

平均评分64

评分/百万令牌1116.2

最佳

GLM 5.2 (free)

最差

LFM2.5-2.6B (free)

Large146 models

平均评分68

评分/百万令牌321.8

最佳

GPT-5.2 Chat

最差

Grok Latest

Mega140 models

平均评分70

评分/百万令牌66.7

最佳

Claude Fable 5.1

最差

GLM Flash Latest

边际递减分析

更大的上下文窗口是否与更高的评分相关？

层级	平均上下文	平均评分	平均效率
Small	9K	52	5808.9
Medium	63K	64	1116.2
Large	240K	68	321.8
Mega	1.1M	70	66.7

输出令牌效率

按输出效率（每1K输出令牌评分）排名的前20个模型。16K+输出令牌的模型已高亮显示。

模型	提供商	评分	最大输出	输出效率
Gemma 2 27BGoogle	Google	77	2K	37.8
MiniMax M2-herMiniMax	MiniMax	68	2K	33.3
GPT-4o (2024-05-13)OpenAI	OpenAI	71	4K	17.4
GPT-4 TurboOpenAI	OpenAI	67	4K	16.3
GPT-4OpenAI	OpenAI	65	4K	15.8
GPT-4 Turbo (batch)OpenAI	OpenAI	62	4K	15.0
Claude 3 HaikuAnthropic	Anthropic	51	4K	12.5
Command R (08-2024)Cohere	Cohere	49	4K	12.2
Command R+ (08-2024)Cohere	Cohere	49	4K	12.2
Schematron V2 Smallinference-net	inference-net	40	4K	9.8
Hy-MT2-1.8BTencent	Tencent	40	4K	9.8
Hy-MT2-30B-A3BTencent	Tencent	40	4K	9.8
Hy-MT2-7BTencent	Tencent	40	4K	9.8
Qwen3 8BAlibaba	Alibaba	61	8K	7.4
Qwen3 235B A22BAlibaba	Alibaba	54	8K	6.6
Command ACohere	Cohere	51	8K	6.2
R1 Distill Llama 70BDeepSeek	DeepSeek	41	7K	5.5
Gemma 4 31BGoogle16K+	Google	81	16K	4.9
Schematron V2 Turboinference-net	inference-net	40	8K	4.9
LFM2.5-2.6B (free)Liquid AI	Liquid AI	40	8K	4.9

关键洞察

从效率数据中自动生成的观察结果。

上下文最优点

Small models have the highest average efficiency at 5808.9 score/MToken across 7 models.

输出很重要

Models with 16K+ output tokens score 30% higher on average than models with smaller output limits.

紧凑型高性能模型

0 models achieve top-20 scores with under 128K context.

探索更多

深入了解上下文窗口、对比模型或探索其他维度。

全部探索器上下文窗口大上下文模型

Frequently Asked Questions

Efficiency is measured as the score-per-context-token ratio - how much ranking score a model achieves relative to its context window size. Models that score highly with smaller context windows are considered more efficient than those requiring massive context to achieve similar results.

Cost efficiency combines quality (composite score) with pricing. The most cost-efficient models achieve high benchmark scores while maintaining low per-token API costs. Free and budget-tier models that perform well are the most cost-efficient options.

Not necessarily. Our efficiency analysis shows diminishing returns beyond certain context sizes. Models with 128K tokens often score similarly to those with 1M+ tokens, meaning the extra context capacity adds cost without proportional quality gains for most use cases.