Model Arena

AI Model Intelligence Hub

模型排名与选型看板

把国际 Arena、国内模型分区、API 定价和发布时间线放到同一页，方便你快速做模型选型、竞品观察和技术跟踪。

数据更新：2026/4/13 12:51:32 · 国际 Arena 来源 LMArena Leaderboard · 时间线参考 AI Flash Report

查看国际榜单查看国内分区比较 API 价格

国际 Arena 榜单

Text / Code / Vision

国内模型分区

覆盖通用到多模态

20+

API 价格样本

便于成本对比

发布时间线

追踪模型演进

国际 Arena 榜单

保留国际主流评测口径，适合看当前头部模型在对话、代码和视觉理解上的整体位置。

查看数据来源

综合对话 Top 1

claude-opus-4-6-thinking

Anthropic · Rating 1503.8

代码 Top 1

claude-opus-4-6-thinking

Anthropic · Rating 1548.4

视觉 Top 1

claude-opus-4-6-thinking

Anthropic · Rating 1302.5

综合对话

Text

来源

#	模型	Rating	组织	投票
1	claude-opus-4-6-thinking	1503.8	Anthropic	16,278
2	claude-opus-4-6	1495.6	Anthropic	17,416
3	muse-spark	1492.8	Meta	3,268
4	gemini-3.1-pro-preview	1491.8	Google	20,531
5	gemini-3-pro	1486.0	Google	41,585
6	grok-4.20-beta1	1485.6	xAI	9,689
7	gpt-5.4-high	1483.9	OpenAI	9,681
8	grok-4.20-beta-0309-reasoning	1478.3	xAI	9,781
9	gpt-5.2-chat-latest-20260210	1476.5	OpenAI	15,704
10	grok-4.20-multi-agent-beta-0309	1476.1	xAI	10,112

代码 / Web 开发

Code

来源

#	模型	Rating	组织	投票
1	claude-opus-4-6-thinking	1548.4	Anthropic	4,015
2	claude-opus-4-6	1542.2	Anthropic	4,841
3	glm-5.1	1529.9	Z.ai	1,046
4	claude-sonnet-4-6	1521.3	Anthropic	6,979
5	claude-opus-4-5-20251101-thinking-32k	1490.2	Anthropic	13,065
6	claude-opus-4-5-20251101	1465.9	Anthropic	14,517
7	gpt-5.4-high (codex-harness)	1456.6	OpenAI	1,485
8	gemini-3.1-pro-preview	1456.3	Google	5,819
9	qwen3.6-plus-preview	1453.3	Alibaba	2,112
10	glm-4.7	1439.4	Z.ai	4,878

视觉理解

Vision

来源

#	模型	Rating	组织	投票
1	claude-opus-4-6-thinking	1302.5	Anthropic	2,388
2	muse-spark	1292.8	Meta	1,210
3	claude-opus-4-6	1288.7	Anthropic	2,525
4	gemini-3-pro	1287.8	Google	13,390
5	gemini-3.1-pro-preview	1278.4	Google	10,116
6	gpt-5.2-chat-latest-20260210	1277.5	OpenAI	7,001
7	claude-sonnet-4-6	1274.2	Anthropic	2,836
8	gemini-3-flash	1267.9	Google	16,184
9	gemini-3-flash (thinking-minimal)	1257.5	Google	14,242
10	dola-seed-2.0-preview	1257.3	Bytedance	4,640

国内模型分区榜单

国内模型分区榜单为站内整理版，综合公开榜单、产品能力、生态落地与近期发布节奏，适合做选型参考，不等同于单一统一基准测试。

更新于 2026/4/13 12:51:32

查看总来源

通用代码 TTS ASR 视频生成图片生成

通用

综合对话、推理、长上下文与产品完成度

Top 5

国内榜单依据

DeepSeek V3.2 DeepSeek

通用对话、推理、长上下文、性价比

MMLU 90.1%，HumanEval 92.5%，1M+ context

推荐

GLM-5 Zhipu AI

综合智能、低幻觉率、国产算力适配

HLE 50.4%，Hallucination Rate 1.2%

Kimi K2 Moonshot AI

长文理解、中文体验、开放权重影响力

LMSYS Arena #1 open-weight，1.04T params

Doubao Seed 2.0 ByteDance

产品化、Agent 场景、多模态联动

多模态能力强，适合字节生态产品化

MiniMax M2.7 MiniMax

通用能力平衡、生成式交互、多模态协同

Arena code 1445，综合产品成熟度高

代码

编码、Agent、工具调用与工程落地能力

Top 5

国内编码榜依据

GLM-5 Zhipu AI

工程代码生成、Agent 编排、中文开发支持

Arena Code Top domestic，HLE 50.4%

推荐

MiniMax M2.7 MiniMax

代码补全、复杂任务拆解、工具调用

Arena Code 1445

GLM-4.7 Zhipu AI

稳定编码、函数调用、工程问答

Arena Code 1439.1

DeepSeek Coder V3 DeepSeek

代码生成、重构、开源生态认可度高

HumanEval / Repo-level coding 表现强

Kimi K2 Moonshot AI

长上下文代码理解、文档到实现链路

长文代码库理解表现突出

TTS

语音合成自然度、情感表现与商用成熟度

Top 5

国内 TTS 榜依据

MiniMax Speech-02 MiniMax

自然度高、情感表达、商业落地成熟

中文自然度与角色语音表现领先

推荐

CosyVoice 2 FunAudioLLM / 阿里系生态

开源可控、零样本音色克隆、中文效果好

开源中文 TTS 代表方案

Step-Audio TTS StepFun

对话式语音、拟人化表达、端到端体验好

语音交互体验强

Doubao Voice ByteDance

产品集成强、延迟低、适合陪伴/内容场景

大规模应用落地能力强

Tencent Cloud TTS Tencent

稳定性高、企业服务成熟、音库丰富

企业级接入成熟

ASR

语音识别准确率、实时性与行业适配

Top 5

国内 ASR 榜依据

Paraformer Large FunAudioLLM / ModelScope

中文识别准确率高、流式与非流式都成熟

中文 ASR 开源标杆

推荐

SenseVoice FunAudioLLM

多语言识别、情感/事件理解、实时性好

ASR + speech understanding 一体化

Tencent Cloud ASR Tencent

稳定、工程接入成熟、行业方案丰富

企业级落地广泛

iFLYTEK Spark ASR iFLYTEK

中文语音识别积累深、行业词表能力强

政企和教育场景强势

Baidu Speech ASR Baidu

云服务稳定、普通话识别成熟、接入门槛低

通用云语音场景覆盖广

视频生成

镜头稳定性、动作表现与一致性

Top 5

国内视频榜依据

Seedance 2.0 ByteDance

音视频同步生成、镜头一致性、商业化能力强

同步音视频单次生成，Languages 8+

推荐

Kling 2.0 Kuaishou

运动幅度、镜头语言、人物动作细节

国内视频生成头部产品

Vidu Q1 ShengShu AI

叙事连贯性、风格控制、中文提示词友好

创作者社区反馈稳定

Wan 2.1 Alibaba

开源路线、可控生成、生态联动强

开源视频模型代表之一

Hailuo Video MiniMax

人物演绎、短视频生成、产品体验完整

消费级生成体验优秀

图片生成

中文提示词理解、审美质量与可控性

Top 5

国内图片榜依据

FLUX China / 国内优化版生态 Open ecosystem

中文提示词适配、写实质感、社区活跃

国内创作者生态采用广泛

推荐

Kolors Kuaishou

中文语义理解强、海报和人物图表现好

中文文生图代表模型

Tongyi Wanxiang Alibaba

电商和设计场景适配、企业服务能力强

企业图像生成落地成熟

Doubao Image ByteDance

社媒内容生成、风格化、上手门槛低

内容创作场景增长快

Ernie Image Baidu

企业集成、通用图像生成、中文理解稳定

云产品体系完整

API 定价对比

统一按每百万 tokens 展示，方便快速看输入、输出与上下文成本结构。

查看数据来源

筛选思路先看输入成本，再看输出倍率和上下文窗口，避免只看单价忽略真实任务成本。

模型	提供商	输入	输出	上下文	最大输出
GPT-4.1 Nano	OpenAI	$0.10	$0.40	1M	16K
Qwen 3.5 Flash	Alibaba	$0.10	$0.40	1M	8K
Gemini 2.5 Flash	Google	$0.15	$0.60	1M	64K
Llama 4 Maverick	Meta (via API)	$0.20	$0.60	1M	16K
GPT-5 Mini	OpenAI	$0.25	$2.00	128K	16K
DeepSeek V3.2	DeepSeek	$0.25	$0.40	128K	16K
Grok 3 Mini	xAI	$0.30	$0.50	128K	16K
GPT-4.1 Mini	OpenAI	$0.40	$1.60	1M	16K
DeepSeek R1	DeepSeek	$0.55	$2.19	128K	16K
Claude Haiku 3.5	Anthropic	$0.80	$4.00	200K	8K
o4-mini	OpenAI	$1.10	$4.40	200K	100K
GPT-5	OpenAI	$1.25	$10.00	128K	16K
Gemini 3.1 Pro	Google	$1.25	$10.00	1M	64K
Gemini 2.5 Pro	Google	$1.25	$10.00	1M	64K
GPT-5.2	OpenAI	$1.75	$14.00	400K	32K
GPT-4.1	OpenAI	$2.00	$8.00	1M	32K
o3	OpenAI	$2.00	$8.00	200K	100K
Claude Sonnet 4.6	Anthropic	$3.00	$15.00	200K	64K
Grok 3	xAI	$3.00	$15.00	128K	16K
Claude Opus 4.6	Anthropic	$5.00	$25.00	200K	32K

模型发布时间线

按时间回看模型迭代节奏，判断各厂商当前重点押注方向。

查看数据来源

观察重点留意各家在代码、推理、多模态和长上下文上的发布时间密度，能更快看出产品路线。

2026-02-19

Gemini 3.1 Pro Google LLM

ARC-AGI-2 77.1%，MMLU 93.8%

2x reasoning improvement
ARC-AGI-2 score of 77.1%
Enhanced multimodal understanding
Deep Think mode

2026-02-17

Claude Sonnet 4.6 Anthropic LLM

SWE-bench 80.8%，MMLU 92.1%

Agent Teams: orchestrate 2-16 Claude instances
Near-Opus performance at 1/5th cost
80.8% SWE-bench Verified
Fast mode research preview

2026-02-12

DeepSeek V3.2 DeepSeek LLM

MMLU 90.1%，HumanEval 92.5%

1M+ token context window (10x expansion)
Improved reasoning capabilities
Open source release
Cost-effective inference

2026-02-11

GLM-5 Zhipu AI LLM

HLE 50.4%，Hallucination Rate 1.2%

First frontier model trained on Huawei Ascend chips (no NVIDIA)
#1 HLE score (50.4%)
1.2% hallucination rate via Slime RL
136x cheaper than Claude Opus 4.5

2026-02-10

Seedance 2.0 ByteDance Video

Audio Sync Excellent，Languages 8+

Synchronized audio + video generation in one pass
Lip-sync in 8+ languages
10-30x cheaper than Sora 2
Commercial pricing: $0.10-$0.80/min

2026-02-05

GPT-5.3 Codex OpenAI Code

Terminal-Bench 77.3%，SWE-Bench Pro SOTA

Self-improving agentic coding
25% faster than GPT-5.2-Codex
1,000+ tokens/sec generation
First OpenAI model flagged 'high' on cybersecurity framework

2026-01-20

Kimi K2 Moonshot AI LLM

LMSYS Arena #1，Parameters 1.04T

First open-weight model #1 on LMSYS Chatbot Arena
1.04 trillion parameters
K2.5 agent swarms with up to 100 sub-agents
$0.15/M input tokens

2025-12-18

GPT-5.2 Codex OpenAI Code

SWE-Bench SOTA，HumanEval 95.1%

Specialized for software engineering
Enhanced agentic coding
Multi-file refactoring
Advanced debugging capabilities

2025-12-15

Mistral Large 3 Mistral LLM

MMLU 89.4%，HumanEval 91.2%

128K context window
Improved multilingual capabilities
Enhanced function calling
Competitive with GPT-5 class models

2025-12-11

GPT-5.2 OpenAI LLM

MMLU 92.8%，MATH 88.5%

Enhanced reasoning capabilities
Improved adaptive reasoning
Better multimodal understanding
Faster inference

2025-11-24

Claude Opus 4.5 Anthropic LLM

SWE-bench 80.9%，MMLU 92.8%

First model to break 80.9% on SWE-Bench Verified
67% price reduction vs previous Opus
Extended reasoning capabilities
Advanced coding performance

2025-11-18

Gemini 3 Pro Google Multimodal

ARC-AGI 87.5%，MMLU 93.2%

1M token context window
Deep Think reasoning mode
Solved 5/6 IMO 2025 problems
#1 on LMSYS Arena

2025-11-12

GPT-5.1 OpenAI LLM

ARC-AGI 87.5%，AIME 2025 100%

Adaptive reasoning modes
Perfect 100% on AIME 2025
87.5% on ARC-AGI
Enhanced multimodal capabilities

2025-08-15

GPT-5 OpenAI LLM

MMLU 91.0%，HumanEval 93.5%

Adaptive reasoning (routes between quick and deep thinking)
Improved math and coding
Enhanced multimodal reasoning
New safety architecture

2025-07-15

Claude Opus 4.1 Anthropic LLM

SWE-bench 75.2%，MMLU 91.2%

Improved multi-file refactoring
Enhanced agentic capabilities
Better long-context performance
Reduced hallucinations

2025-06-20

Gemini 2.5 Flash Google Multimodal

MMLU 87.5%，Speed 2x Gemini 2.0 Flash

Enhanced image editing stabilization
Faster inference
Improved multimodal understanding
Cost-effective deployment

2025-05-22

Claude Sonnet 4 Anthropic LLM

MMLU 88.7%，HumanEval 94.5%

Enhanced reasoning capabilities
Improved safety measures
Advanced multimodal understanding
Extended context window

2025-02-24

Claude Sonnet 3.7 Anthropic LLM

MMLU 86.1%，HumanEval 93.2%

Improved reasoning
Better code generation
Enhanced safety
Reduced hallucinations

2024-12-26

DeepSeek-V3 DeepSeek LLM

MMLU 88.5%，HumanEval 82.6%

Mixture of Experts architecture
Cost-effective training
Open source release
Strong reasoning capabilities

2024-12-11

Gemini 2.0 Flash Google Multimodal

MMLU 85.8%，HumanEval 71.9%

Native multimodal generation
Real-time API
Agentic capabilities
Enhanced speed

2024-10-16

Moonshot Kimi Moonshot AI LLM

C-Eval 77.9%，CMMLU 75.2%

200K context window
Multilingual support
Document processing
Chinese language optimization

2024-09-12

GPT-o1-preview OpenAI LLM

AIME 83rd percentile，GPQA 78%

Advanced reasoning capabilities
Chain-of-thought processing
Enhanced problem-solving
Improved mathematical reasoning

2024-08-13

Grok-2 xAI LLM

MMLU 84.0%，HumanEval 74.1%

Real-time information access
Multimodal understanding
X platform integration
Conversational AI

2024-07-25

GitHub Copilot Chat Microsoft Code

HumanEval 84.2%，MBPP 78.9%

Conversational coding
IDE integration
Code explanation
Multi-language support

2024-06-20

Claude 3.5 Sonnet Anthropic LLM

MMLU 88.7%，HumanEval 92.0%

200K context window
Improved coding capabilities
Enhanced reasoning
Vision capabilities

2024-03-28

Grok-1.5 xAI LLM

MMLU 73.0%，HumanEval 63.2%

Improved reasoning
Code generation
128K context window
Real-time data access

2024-03-04

Claude 3 Opus Anthropic LLM

MMLU 86.8%，HumanEval 84.9%

200K context window
Advanced reasoning
Multimodal capabilities
Constitutional AI training

2024-03-04

Claude 3 Sonnet Anthropic LLM

MMLU 79.0%，HumanEval 73.0%

200K context window
Balanced capability and speed
Multimodal input
Strong reasoning

2024-03-04

Claude 3 Haiku Anthropic LLM

MMLU 75.2%，HumanEval 75.9%

200K context window
Fastest response times
Multimodal input
Cost-effective

2024-02-26

Mistral Large Mistral LLM

Top-tier reasoning model with strong multilingual capabilities。MMLU 81.2%，HumanEval 45%

32K context window
Multilingual capabilities
Function calling
JSON mode

2024-02-15

Sora OpenAI Video

Video Quality High，Motion Coherence Excellent

Text-to-video generation
60-second video clips
Complex scene generation
Realistic motion physics

2024-02-15

Gemini 1.5 Pro Google Multimodal

MMLU 81.9%，HumanEval 71.9%

1M token context window
Multimodal understanding
Video analysis
Audio processing

2024-01-25

text-embedding-3-large OpenAI Embedding

MTEB Score 64.6%，Dimensions 3072

3072 embedding dimensions
Improved retrieval performance
Reduced hallucinations
Multi-language support

2024-01-25

GPT-4 Turbo OpenAI LLM

MMLU 86.4%，HumanEval 67%

128K context window
Improved instruction following
Enhanced reasoning capabilities
Reduced hallucinations

2023-12-21

Midjourney v6 Midjourney Vision

Image Quality Very High，Prompt Adherence 92%

Photorealistic image generation
Improved prompt understanding
Better human anatomy
Text rendering

2023-12-07

Grok-1 xAI LLM

MMLU 73.0%，HumanEval 63.2%

Real-time information
Conversational interface
X platform integration
Uncensored responses

2023-12-06

AlphaCode 2 Google DeepMind Code

Codeforces Rating 1747，Problem Solving 85th percentile

Advanced code generation
Competitive programming
Multi-language support
Problem decomposition

2023-12-06

Gemini Ultra Google Multimodal

MMLU 90.0%，HumanEval 74.4%

Multimodal reasoning
Text, image, audio, video understanding
Advanced mathematical reasoning
Code generation

2023-12-06

Gemini Pro Google Multimodal

MMLU 79.1%，HumanEval 67.7%

Multimodal capabilities
32K context window
Fast inference
Scalable deployment

2023-11-21

Claude 2.1 Anthropic LLM

MMLU 73.1%，HumanEval 70.0%

200K context window
Reduced hallucination rates
Enhanced accuracy
Tool use capabilities