上下文窗口效率探索器

分析297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

效率概览

所有分析模型的关键效率指标。

最高效 (128K+)

GPT-5.3 Chat

664.1 score/MToken

最佳输出效率

Inflection 3 Productivity

35.7 score/1K output

最佳成本效率

LFM2-8B-A1B

3546.7 score/$

平均总体效率

7.3%

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	Llemma 7beleutherai	eleutherai	47	4K	4K	11547.9	11.5	47.3	Small
2	GPT-3.5 Turbo (older v0613)OpenAI	OpenAI	38	4K	4K	9279.6	9.3	25.3	Small
3	GPT-3.5 Turbo InstructOpenAI	OpenAI	32	4K	4K	7863.2	7.9	18.4	Small
4	Gemma 2 27BGoogle	Google	60	8K	2K	7263.2	29.1	91.5	Small
5	Gemma 3n 2B (free)Google	Google	58	8K	2K	7080.1	28.3	Free	Small
6	Mistral 7B Instruct v0.1Mistral AI	Mistral AI	20	3K	--	7011.3	--	132.0	Small
7	Gemma 3n 4B (free)Google	Google	55	8K	2K	6750.5	27.0	Free	Small
8	Llama 3 70B InstructMeta	Meta	40	8K	8K	4931.6	5.0	64.6	Small
9	GPT-4 (older v0314)OpenAI	OpenAI	39	8K	4K	4761.3	9.5	0.9	Small
10	GPT-4OpenAI	OpenAI	39	8K	4K	4761.3	9.5	0.9	Small
11	Reka Edgereka	reka	77	16K	16K	4711.9	4.7	386.0	Small
12	Inflection 3 ProductivityInflection	Inflection	37	8K	1K	4575.0	35.7	5.9	Small
13	Inflection 3 PiInflection	Inflection	37	8K	1K	4575.0	35.7	5.9	Small
14	Llama 3 8B InstructMeta	Meta	30	8K	16K	3698.7	1.8	865.7	Small
15	Gemma 2 9BGoogle	Google	30	8K	--	3674.3	--	501.7	Small
16	Phi 4Microsoft	Microsoft	60	16K	16K	3631.6	3.6	580.5	Small
17	Llama 3.1 8B InstructMeta	Meta	42	16K	16K	2575.7	2.6	1205.7	Small
18	ERNIE 4.5 VL 28B A3BBaidu	Baidu	75	30K	8K	2493.3	9.3	213.7	Small
19	GPT-3.5 Turbo 16kOpenAI	OpenAI	40	16K	4K	2435.2	9.7	11.4	Small
20	GPT-3.5 TurboOpenAI	OpenAI	40	16K	4K	2435.2	9.7	39.9	Small
21	DeepSeek V3.1DeepSeek	DeepSeek	74	33K	7K	2246.1	10.3	163.6	Medium
22	Rnj 1 Instructessentialai	essentialai	65	33K	--	1977.5	--	432.0	Medium
23	Gemma 3 4B (free)Google	Google	61	33K	8K	1858.5	7.4	Free	Medium
24	Qwen2.5 VL 72B InstructAlibaba	Alibaba	60	33K	33K	1837.2	1.8	75.3	Medium
25	R1 Distill Qwen 32BDeepSeek	DeepSeek	60	33K	33K	1831.1	1.8	206.9	Medium
26	Mistral Small 3Mistral AI	Mistral AI	59	33K	16K	1809.7	3.6	912.3	Medium
27	LFM2.5-1.2B-Thinking (free)Liquid AI	Liquid AI	59	33K	--	1800.5	--	Free	Medium
28	Mistral Small CreativeMistral AI	Mistral AI	59	33K	--	1800.5	--	295.0	Medium
29	Qwen-Max Alibaba	Alibaba	59	33K	8K	1788.3	7.2	22.5	Medium
30	Qwen3 30B A3BAlibaba	Alibaba	71	41K	41K	1740.7	1.7	396.1	Medium
31	Qwen3 14BAlibaba	Alibaba	71	41K	41K	1738.3	1.7	474.7	Medium
32	Qwen3 32BAlibaba	Alibaba	71	41K	41K	1738.3	1.7	445.0	Medium
33	Gemma 3 12B (free)Google	Google	55	33K	8K	1678.5	6.7	Free	Medium
34	LFM2-24B-A2BLiquid AI	Liquid AI	53	33K	--	1623.5	--	709.3	Medium
35	LFM2.5-1.2B-Instruct (free)Liquid AI	Liquid AI	53	33K	--	1623.5	--	Free	Medium
36	LFM2-8B-A1BLiquid AI	Liquid AI	53	33K	--	1623.5	--	3546.7	Medium
37	LFM2-2.6BLiquid AI	Liquid AI	53	33K	--	1623.5	--	3546.7	Medium
38	SabaMistral AI	Mistral AI	53	33K	--	1608.3	--	131.8	Medium
39	Qwen2.5 72B InstructAlibaba	Alibaba	52	33K	16K	1596.1	3.2	205.1	Medium
40	Qwen3 8BAlibaba	Alibaba	65	41K	8K	1584.5	7.9	288.4	Medium
41	Qwen3 4B (free)Alibaba	Alibaba	63	41K	--	1533.2	--	Free	Medium
42	Gemma 3n 4BGoogle	Google	46	33K	--	1406.9	--	1536.7	Medium
43	Qwen2.5 7B InstructAlibaba	Alibaba	46	33K	33K	1388.5	1.4	650.0	Medium
44	Coder Largearcee-ai	arcee-ai	45	33K	--	1382.4	--	69.7	Medium
45	Qwen2.5 Coder 7B InstructAlibaba	Alibaba	43	33K	--	1306.2	--	713.3	Medium
46	Mixtral 8x7B InstructMistral AI	Mistral AI	42	33K	16K	1293.9	2.6	78.5	Medium
47	Qwen2.5 Coder 32B InstructAlibaba	Alibaba	42	33K	--	1290.9	--	51.0	Medium
48	R1DeepSeek	DeepSeek	68	64K	16K	1065.6	4.3	42.6	Medium
49	Olmo 3 32B ThinkAllen AI	Allen AI	66	66K	66K	1011.7	1.0	204.0	Medium
50	Olmo 3.1 32B InstructAllen AI	Allen AI	65	66K	--	990.3	--	162.3	Medium

层级分析

不同上下文窗口层级的效率分析。

Small20 models

平均评分45

评分/百万令牌5302.8

最佳

Llemma 7b

最差

GPT-3.5 Turbo 16k

Medium41 models

平均评分56

评分/百万令牌1327.7

最佳

DeepSeek V3.1

最差

Sonar

Large198 models

平均评分71

评分/百万令牌413.7

最佳

GPT-5.3 Chat

最差

GPT-5 Nano

Mega38 models

平均评分81

评分/百万令牌74.4

最佳

Claude Opus 4.6

最差

Grok 4.20 Multi-Agent Beta

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	10K	45	5302.8
Medium	51K	56	1327.7
Large	191K	71	413.7
Mega	1.1M	81	74.4

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
Inflection 3 ProductivityInflection	Inflection	37	1K	35.7
Inflection 3 PiInflection	Inflection	37	1K	35.7
UI-TARS 7B ByteDance	ByteDance	63	2K	30.5
Gemma 2 27BGoogle	Google	60	2K	29.1
MiniMax M2-herMiniMax	MiniMax	59	2K	29.0
Gemma 3n 2B (free)Google	Google	58	2K	28.3
Gemma 3n 4B (free)Google	Google	55	2K	27.0
Jamba Large 1.7AI21 Labs	AI21 Labs	71	4K	17.3
GPT-4 TurboOpenAI	OpenAI	60	4K	14.7
GPT-4o (2024-05-13)OpenAI	OpenAI	53	4K	12.8
Command R+ (08-2024)Cohere	Cohere	48	4K	11.9
Command R (08-2024)Cohere	Cohere	48	4K	11.9
Llemma 7beleutherai	eleutherai	47	4K	11.5
Nova Lite 1.0Amazon	Amazon	58	5K	11.3
Nova Pro 1.0Amazon	Amazon	58	5K	11.3
Command R7B (12-2024)Cohere	Cohere	45	4K	11.2
Sonar Pro SearchPerplexity	Perplexity	85	8K	10.6
Claude 3 HaikuAnthropic	Anthropic	43	4K	10.5
GPT-4 Turbo PreviewOpenAI	OpenAI	43	4K	10.4
GPT-4 Turbo (older v1106)OpenAI	OpenAI	43	4K	10.4

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 5302.8 score/MToken across 20 models.

输出很重要

Models with 16K+ output tokens score 35% 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.