AI Models

Compare leading AI models.

Claude 3 Haiku

Anthropic

Anthropic's fastest and cheapest model, ideal for real-time conversations and high-frequency API calls with extremely low cost.

Claude 3.5 Sonnet

Anthropic

Anthropic's strongest coding model, ranked first on SWE-bench, with top-tier code quality and instruction-following capabilities, excelling in agent tasks.

Claude Haiku 4.5

Anthropic

Claude Haiku 4.5, the fastest Claude, with near-frontier intelligence and 200K context, ideal for high-concurrency and low-latency scenarios.

Claude Opus 4.5

Anthropic

Claude Opus 4.5, 200K context, high-quality reasoning and coding, relatively better cost-effectiveness.

Claude Opus 4.6

Anthropic

Claude Opus 4.6, 1M context, supports extended thinking, stable performance on complex tasks.

Claude Opus 4.7

Anthropic

Claude Opus, the previous generation flagship, offers 1M context, strong complex reasoning and agentic coding capabilities, and remains a top-tier choice.

Claude Opus 4.8

Anthropic

Anthropic's current strongest model, top-tier in complex reasoning, long-cycle agentic coding, and highly autonomous tasks, ranked first in the Intelligence Index.

Claude Sonnet 4.5

Anthropic

Claude Sonnet 4.5, with 200K context, balances speed and intelligence, excelling in coding and agent tasks.

Claude Sonnet 4.6

Anthropic

The best balance of speed and intelligence, with 1M context, offering great value for daily development and agent tasks.

Codestral

Mistral

Mistral code-specific model, supporting 80+ programming languages, code completion and generation.

DeepSeek V4 Flash

DeepSeek

DeepSeek's fast open-source model with 1M context and extremely low price, suitable for large-scale batch processing.

DeepSeek V4 Pro

DeepSeek

China's open-source flagship, with 1M context and dual thinking/non-thinking modes, excels in coding and reasoning, priced at a fraction of closed-source flagships.

DeepSeek-R1

DeepSeek

Designed for complex reasoning, with math/logic/coding capabilities comparable to o1, but fully open-source with training costs at only 3%.

DeepSeek-V3

DeepSeek

Domestic flagship model with code and math capabilities comparable to Claude, priced at only 5% of OpenAI, the king of cost-effectiveness.

Gemini 1.5 Pro

Google

Google's ultra-long context specialized model with a 2 million token window, capable of analyzing entire codebases or long videos.

Gemini 2.0 Flash

Google

Google's latest agentic model with million-token ultra-long context, native tool calling support, and extremely low price.

Gemini 3.1 Pro

Google

Google's flagship, with ultra-long context, native multimodal capabilities, and tool calling, its reasoning ability ranks among the top tier.

Gemini 3.5 Flash

Google

Gemini Fast: excellent speed and cost, ideal for high concurrency and real-time scenarios.

GLM-4-Plus

Zhipu

Zhipu GLM-4-Plus: Closed-source flagship version (2024) with strong overall capabilities.

GLM-4.5

Zhipu

Zhipu GLM-4.5: 128K context, comprehensive capabilities comparable to mainstream models, one of the leading domestic open-source models.

GLM-4.6

Zhipu

Zhipu GLM-4.6: 200K context (expanded from 128K in 4.5), token efficiency improved by approximately 30% compared to the previous generation.

GLM-4.7

Zhipu

Zhipu GLM-4.7: Open source (MIT), an iterative upgrade of GLM-4.6.

GLM-5

Zhipu

Zhipu GLM-5: The latest open-source flagship in 2026 (MIT), with comprehensive capabilities comparable to mainstream closed-source models.

GPT-4o

OpenAI

OpenAI's flagship multimodal model, proficient in vision, speech, and text, with fast response speed and the most comprehensive ecosystem.

GPT-4o mini

OpenAI

GPT-4o Lite, 3x faster than GPT-4o with 95% cost reduction, ideal for high-concurrency agent scenarios.

GPT-5.5

OpenAI

OpenAI's current flagship, ranked second on the Intelligence Index (just behind Claude Opus 4.8), with top-tier performance in high/ultra-high reasoning settings.

Kimi K2

Moonshot

Kimi K2: 128K–256K context, 1T total params (32B active) MoE, trained on 15.5T tokens, open-source weights.

Kimi K2 Thinking

Moonshot

Kimi K2 Thinking: 256K context, 1T parameters (32B activated) MoE, focused on deep reasoning, open-source weights.

Kimi K2.5

Moonshot

Moonshot Kimi K2.5: 256K context, 1T total params (32B active) MoE + vision, open-source weights, multimodal upgrade of K2.

Kimi K2.6

Moonshot

Moonshot Kimi series, one of the highest-ranked open-source weight models on the current Intelligence Index.

Llama 3.1 405B

Meta

Meta Llama 3.1 405B: A flagship open-source model with 405 billion parameters, 128K context, and performance approaching closed-source flagships.

Llama 3.1 70B

Meta

Meta Llama 3.1 70B: A 70-billion-parameter open-source general-purpose model with 128K context, offering high deployment cost-effectiveness.

Llama 3.1 8B

Meta

Meta Llama 3.1 8B: An 8-billion-parameter lightweight open-source model with 128K context, suitable for edge and low-cost deployment.

Llama 3.3 70B

Meta

Meta's latest open-source flagship, 70 billion parameters, self-hostable, business-friendly license, performance approaching closed-source models.

Llama 4 Maverick

Meta

Meta Llama 4 Maverick: 400B total params (17B activated, 128 experts) MoE, 1M context, open weights, commercially friendly.

Llama 4 Scout

Meta

Meta's latest open-source model features a native 10M ultra-long context, supports self-hosting, and is commercially friendly.

Mistral Large

Mistral

Europe's strongest AI model, with excellent multilingual capabilities and support for Function Calling, suitable for European compliance scenarios.

Mistral Large 2

Mistral

Mistral flagship, 128K context, strong multilingual and code capabilities, representative of European open-weight models.

Mixtral 8x22B

Mistral

Mistral's flagship MoE architecture open-source model, with 141 billion parameters and 39 billion activated, achieving the best balance between performance and cost.

Mixtral 8x7B

Mistral

Mistral classic MoE model, 8×7B experts, open-source, lightweight and efficient, widely deployed.

o1-preview

OpenAI

OpenAI's specialized reasoning model, answers after deep thinking, performs best on complex math/science/code problems, but is slow.

o3

OpenAI

OpenAI o-series reasoning model, answers after deep thinking, performs strongly on complex math/science/coding problems.

Qwen2.5-72B

Alibaba

Alibaba's Tongyi Qianwen latest flagship, with the strongest Chinese language capabilities domestically, fully open-source, and supports multimodal.

Qwen2.5-Coder

Alibaba

Alibaba's specialized code model surpasses Claude 3.5 Sonnet in coding ability, achieving 98.5% on HumanEval, fully open-source.

Qwen2.5-Max

Alibaba

Tongyi Qianwen 2.5 Max: A large-scale MoE flagship model with comprehensive capabilities comparable to mainstream closed-source models.

Qwen3-Coder

Alibaba

Tongyi Qianwen 3 Code Special: 480B-A35B MoE open source (Apache 2.0), strong code capabilities.

Qwen3-Max

Alibaba

Tongyi Qianwen 3 Max: A closed-source flagship with over 1T parameters, representing the ceiling of the Tongyi series' capabilities.

Qwen3.5

Alibaba

Alibaba Tongyi Qianwen 3.5, open-source, fast, and extremely low-cost (starting from approximately $0.01/1M), available in multiple sizes.

Qwen3.6

Alibaba

Tongyi Qianwen 3.6: 35B-A3B MoE open-source model (Apache 2.0), the latest generation in 2026.

Qwen3.6-Plus

Alibaba

Tongyi Qianwen 3.6 Plus: Closed-source flagship version, released in 2026, with comprehensive capabilities comparable to mainstream closed-source models.

QwQ-32B

Alibaba

Tongyi Qianwen specialized open-source reasoning model with 32 billion parameters, excelling in mathematics and logical reasoning.

Step-2

StepFun

Step-2 by StepStar: a trillion-parameter LLM (2024), one of the representative domestic large models.

Step-3

StepFun

Step-3 by StepFun: A large model released in 2025, focusing on multimodality and large-scale parameters.

Step-3.5-Flash

StepFun

Step-3.5-Flash by Jieyue Xingchen: 196B-A11B MoE, 256K context, supports tool calling, open-source (Apache 2.0).

Step-3.7-Flash

StepFun

Step-3.7-Flash by Jieyue Xingchen: The latest open-source (Apache 2.0) fast model in 2026.