Available Models

Claude Haiku 4.5 is Anthropic’s fastest and most cost-efficient model, delivering performance comparable to Sonnet 4 across coding, computer use, and agentic tasks.

Claude Opus 4.5 is Anthropic’s frontier reasoning model, designed for complex software engineering, agentic workflows, and long-running computer tasks. It delivers strong multimodal performance, excels on coding and reasoning benchmarks, and offers improved resistance to prompt injection attacks.The model allows developers to adjust inference effort based on task needs, balancing response speed, reasoning depth, and token usage. It also supports advanced tool use, extended context management, and multi-agent workflows, making it well suited for research, code debugging, planning, and browser or spreadsheet tasks.

Anthropic has released the new-generation Claude Opus 4.6 model, featuring a 1M context window and increasing the maximum output token limit to 128K—double the previous generation’s 64K cap. The model introduces an adaptive thinking mode that dynamically adjusts reasoning depth based on task complexity, along with a new highest-level “max effort” parameter.

Opus 4.7 has made significant improvements over Opus 4.6 in advanced software engineering, showing clear progress on the most challenging tasks. It can confidently handle the most complex coding work—which previously required close supervision. Opus 4.7 processes complex and long-running tasks in a rigorous and consistent manner, follows instructions precisely, and devises methods to verify its output before reporting.

Claude Sonnet 4.5 is the world’s leading coding model and Anthropic’s most capable model for building complex agents and using computers. It also delivers significant improvements in reasoning and mathematics.

Sonnet 4.6 is Anthropic’s most powerful Sonnet model to date, achieving frontier-level performance in programming, AI agents, and professional work. It excels at iterative development, navigating complex codebases, end-to-end project management with memory, professional document writing, and reliable computer use for web-based question answering and workflow automation.

Claude Sonnet 5 is built to be the most dominant Sonnet model to date. It can formulate plans, use tools such as browsers and terminals, and operate autonomously, reaching a level that just a few months ago would have required a much larger and more expensive model. For many developers, the era of agentic AI began with Sonnet-class models: Claude Sonnet 3.5,3.6 and 3.7 were the first to demonstrate astonishing capabilities in coding and tool use. Recently, however, the most notable advancement in agentic capabilities has emerged in our Opus-tier models.

DeepSeek-V4-Pro is a high-performance open-source large model released by DeepSeek. It features top-tier reasoning and Agent capabilities, supports ultra-long context, is compatible with domestic Ascend chips, and offers exceptional cost-effectiveness.

DeepSeek-V4-Flash is the lightweight version of the DeepSeek V4 series, focusing on high cost-effectiveness and high throughput efficiency. It is suitable for general dialogue and basic text tasks, while supporting a million-Token long context and efficient inference.

Seedance 2.0 is a new generation of professional-grade multimodal video creation model launched by the Doubao Model Team. It supports multiple modalities such as images, videos, and audio as reference inputs to generate videos, breaking the limitations of creating with single materials. It also has video editing and video extension capabilities, and can accurately reproduce the details, materials, timbre, visual effects style, and camera movement of objects. Character characteristics can also be stably maintained, giving creators control like a director.

Seedance 2.0 Fast supports generating videos from multimodal materials such as images, videos, and audio, while also possessing video editing and extension capabilities, enabling video generation tools to enter a new industrialized stage of accurate generation and reusable iteration. The model's understanding of physical laws continues to deepen, becoming more closely aligned with the real world, and its intent understanding ability is significantly improved, strictly adhering to the constraints of instruction details, thereby ensuring the credibility of professional-grade narratives.

GLM-4.7 is Zhipu’s latest flagship model. It is enhanced for Agentic Coding scenarios, with stronger coding ability, long-horizon task planning, and tool coordination, and it has achieved leading performance among open-source models on current rankings across multiple public benchmarks. Its general capabilities have also been improved, with more concise and natural responses and more immersive writing.

GLM-5 is Zhipu’s next-generation flagship foundation model, built for Agentic Engineering and designed to deliver reliable productivity in complex systems engineering and long-horizon agent tasks. In coding and agent capabilities, GLM-5 achieves state-of-the-art performance among open-source models. In real-world programming scenarios, its user experience approaches that of Claude Opus 4.5. It excels at complex systems engineering and long-horizon agent tasks, making it an ideal foundation for general-purpose agent assistants.

GLM-5.2 is a flagship model designed for the era of long tasks. It supports a truly usable 1M context, and in actual tests, it can handle project-level engineering context, enabling more stable execution of long-range tasks, more reliable adherence to engineering specifications, and further improvement in the success rate of development scenarios. A single task can complete the entire development chain from requirements to multi-terminal deployable products.

glm-5.1 is the latest flagship model launched by the Zhipu platform.

GPT-5.1-Codex-Max is built on an updated version of the base reasoning model, which has been trained to handle agent tasks in fields such as software engineering, mathematics, research, medicine, and computer applications. This is our first natively trained model that supports multiple context windows, capable of coherently processing millions of tokens in a single task through the process of compaction. Like its predecessor, GPT-5.1-Codex-Max has been trained on real-world software engineering tasks, such as Pull Request (PR) creation, code review, front-end development, and question answering.

GPT-5.1-Codex mini is a lightweight, cost-efficient variant of GPT-5.1-Codex optimized for coding, software development, and interactive programming tasks.

Gpt-5.4-Nano is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

GPT Image 1.5 is our latest image generation model, with better instruction tracking and adherence to prompts.

GPT Image 2 is a state-of-the-art image generation model that supports fast, high-quality image generation and editing. It supports flexible image sizes and high-fidelity image input.

GPT-4.1 is a high-performance multimodal model released by OpenAI on April 15, 2025, positioned as a comprehensive upgrade to GPT-4o. It excels in advanced code generation, long-context understanding, fast inference, and multimodal processing, with key strengths of a 1M-token context window and lower cost, optimized for software development, complex instruction execution, and long-document analysis.

GPT-4.1 Mini is a lightweight multimodal large language model launched by OpenAI. It balances strong reasoning capability, million-level ultra-long context window and ultra-low cost, with low latency and high throughput. It is suitable for large-scale high-concurrency services, long document processing, daily conversation and lightweight multimodal application scenarios.

GPT-4.1 Nano is the smallest and most cost-efficient model in OpenAI’s GPT-4.1 family. It is designed for very low-cost, high-throughput tasks such as classification, extraction, simple reasoning, lightweight coding assistance, data transformation, and large-scale automation. It supports long-context text processing and image understanding, but its reasoning and coding ability are generally weaker than GPT-4.1 Mini and the full GPT-4.1 model.

Gpt-5 is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

GPT-5 Chat Latest is a new flagship multimodal conversational large model launched by OpenAI. It delivers powerful logical reasoning, ultra-low hallucinations, ultra-long context comprehension and excellent multimodal analysis capabilities, balancing response speed with professional-level generation quality. It is suitable for daily conversation, complex problem solving, long document analysis, creative content creation and multimodal intelligent interaction scenarios.

GPT-5-Codex is a professional large model released by OpenAI on September 15, 2025, deeply optimized for agentic coding and software engineering based on GPT-5OpenAI. Trained on real-world development scenarios, it delivers both fast interactive responses and long-duration independent task execution (over 7 hours of continuous work). It excels in code generation, review, debugging, refactoring, multi-language development, and system-level interaction, deeply integrated with IDE, CLI, cloud, and GitHub workflows—serving as an AI coding teammate for professional developers.

GPT-5 Mini is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

Gpt-5-Nano is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

Gpt-5-Pro is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

Gpt-5.1 is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

Gpt-5.1-Chat-Latest is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

GPT-5.2 is the best model for coding and intelligence tasks across industries.

Gpt-5.2-Chat-Latest is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

Gpt-5.2-Codex is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

Gpt-5.2-Pro is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

GPT-5.3-Codex redefines the role of AI in programming and general productivity through improved performance, broader capability generalization, and enhanced safety.

Small and affordablGPT-5.4 is our frontier model for complex professional work.e model for lightweight tasks

GPT-5.4 Pro uses more compute to think more deeply and deliver consistently better answers. It is available exclusively through the Responses API, which supports multi-turn model interaction before generating a response and will enable additional advanced API capabilities in the future.

GPT-5.5 (codenamed "Spud") is OpenAI's flagship multimodal large model released on April 23, 2026. As the first complete ground-up retrain since GPT-4.5, it is positioned as an agentic model for real work, featuring native computer control, autonomous agents, 1M+ token context, deep reasoning, and advanced coding & scientific capabilities, optimized for complex professional workflows, autonomous programming, long-document analysis, and enterprise AI applications.

Gemini 2.5 Flash is a multimodal large language model launched by Google DeepMind in April 2025. It features ultra-fast response, million-level context window, controllable reasoning (Thinking Budget) and high cost performance, serving as the flagship version of the Gemini 2.5 family for large-scale real-time applications.

Gemini 2.5 Flash-Lite is an ultra-lightweight multimodal large model launched by Google DeepMind in June 2025. It features ultra-low cost, ultra-low latency, and high throughput, supports a 1M-token context window and controllable reasoning (Thinking Budget), serving as the cost-effective flagship of the Gemini 2.5 family for large-scale, high-concurrency, lightweight real-time applications (e.g., classification, translation, data processing).

Gemini 2.5 Pro is Google’s latest AI model and its most advanced model to date, excelling at coding and complex prompts. With “Deep Think,” it can reason before responding, improving performance and accuracy. The model performs exceptionally well across multiple benchmarks and ranks first on the LMArena leaderboard for reasoning and code generation. It supports multimodal input, including text, images, audio, video, and code.

Gemini 3 Flash Preview is a preview multimodal large model released by Google DeepMind in December 2025. It features ultra-fast inference, near-Pro level intelligence, 1M-token context window, multimodal inputs (text/images/audio/video), and configurable thinking levels, designed specifically for high-concurrency real-time scenarios such as agentic workflows, interactive development, long-document analysis, and coding.

Gemini 3 Pro Preview is part of Google’s most intelligent model family to date, built on advanced reasoning capabilities. It is designed to turn ideas into reality through agentic workflows, autonomous coding, and complex multimodal tasks.

Gemini 3.1 is Google’s most intelligent model family to date, built on advanced reasoning capabilities. It is designed to turn ideas into reality through agentic workflows, autonomous coding, and complex multimodal tasks. Gemini 3.1 Pro Preview is best suited for complex tasks requiring broad world knowledge and advanced cross-modal reasoning.

GLM-5-Turbo is a foundation model deeply optimized for OpenClaw’s Lobster scenarios. From the training stage, it has been specially optimized around the core requirements of Lobster tasks, strengthening key capabilities such as tool calling, instruction following, scheduled and persistent task handling, and long-chain execution. This enables it to be truly actionable even in complex, dynamic, and long-horizon tasks.

The Grok 4 model possesses deep reasoning capabilities and is trained on xAI's Colossus supercomputer, promising stronger logical reasoning and text generation abilities. xAI claims it is the world's most powerful AI model, achieving PhD-level performance in handling academic problems. It excels in real-time speed, reasoning ability, and advanced vision.

Grok-4-Fast-Non-Reasoning is an xAI Grok-family model aimed at conversational reasoning, real-time style assistance, coding, and agentic workflows, with fast and reasoning variants optimized for different latency and depth trade-offs.

Grok-4-Fast-Reasoning is an xAI Grok-family model aimed at conversational reasoning, real-time style assistance, coding, and agentic workflows, with fast and reasoning variants optimized for different latency and depth trade-offs.

Grok-4-1-fast-non-reasoning is an AI model developed by xAI, optimized for maximum speed in generating responses and carrying out agentic tasks. Unlike its “reasoning” counterpart, this variant skips the use of “thinking tokens,” allowing it to deliver immediate, pattern-matched answers for simple and straightforward queries.

Grok 4.1 excels in creative, emotional, and collaborative interactions, showing a sharper ability to capture subtle intent and delivering a more engaging conversational experience. It also maintains a highly consistent personality while fully inheriting the sharp intelligence and reliable performance of its predecessor.

Grok-Code-Fast-1 is an xAI Grok-family model aimed at conversational reasoning, real-time style assistance, coding, and agentic workflows, with fast and reasoning variants optimized for different latency and depth trade-offs.

Kimi K2.5 is Kimi’s smartest model to date, achieving open-source state-of-the-art performance in agents, coding, visual understanding, and a wide range of general intelligence tasks. It is also Kimi’s most versatile model so far, featuring a natively multimodal architecture that supports both visual and text inputs, thinking and non-thinking modes, as well as dialogue and agent tasks.

Kimi-K2.6 is a Moonshot AI Kimi model designed for long-context understanding, multilingual chat, coding, document analysis, and agentic tasks, with newer K2 variants emphasizing stronger reasoning and tool-use capability.

Meta's latest mixture-of-experts model

MiMo-V2.5-Pro. It is our most powerful model to date, achieving significant improvements over its predecessor, MiMo-V2-Pro, in terms of overall intelligence capabilities, complex software engineering, and long-term tasks. MiMo-V2.5-Pro is a mixture-of-experts model with 1.02T parameters, 42B active parameters, based on a hybrid attention architecture, and has a 1 million token context window.

MiMo-V2.5 is a sparse MoE (15B activated) model with 310B parameters, trained using 48T tokens. Its language backbone inherits the MiMo-V2-Flash hybrid sliding window attention architecture and is equipped with dedicated visual and audio encoders (both pre-trained internal encoders), connected via a lightweight projector.

MiniMax-M2.1 is a lightweight, cutting-edge large language model optimized for coding, agentic workflows, and modern application development. Activating only 10 billion parameters, it delivers a major leap in real-world capabilities while maintaining excellent latency, scalability, and cost efficiency.

MiniMax-M2.5 has achieved or surpassed the industry state of the art across productivity scenarios such as coding, tool use, search, and office work.

MiniMax-M2.7 is a MiniMax model for multilingual chat, reasoning, coding, and agent workflows, typically positioned as a high-throughput text model with competitive cost and broad enterprise use cases.

M3 achieves state-of-the-art performance in professional tasks such as encoding and proxy work. It adopts the new attention architecture MSA (MiniMax Sparse Attention) proposed by our team and supports an ultra-long context window of up to 1 million tokens. Excitingly, it is also a native MultiModal Machine Learning model that supports image and video input and can operate desktop computers.

Mistral's flagship model for complex tasks

Gemini 2.5 Flash Image (codename "Nano Banana") is an image generation and editing model launched by Google DeepMind in August 2025. It features ultra-fast text-to-image, precise natural language editing, multi-image fusion, character consistency, and real-world physical reasoning, serving as the professional imaging model in the Gemini 2.5 family for creative design, e-commerce, and content creation.

Nano Banana 2 provides high-quality image generation and conversational editing at a mainstream price point and low latency.

Gemini 3 Pro Image Preview (Nano Banana Pro) is the next-generation AI image generation and editing model in Google’s Gemini family, serving as an upgraded version of Gemini 2.5 Flash Image (Nano Banana). Combining a multimodal Transformer with diffusion-based modeling, it natively supports 2K (2048×2048) and 4K output, with significant improvements in image quality, text rendering, and physical reasoning.

O3 is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

O3-Mini is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

O3-Pro is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

O4-Mini is an OpenAI model for general-purpose reasoning, coding, instruction following, tool use, and production chat workloads, with the exact speed, reasoning depth, and cost depending on the selected variant.

The QwQ inference model trained on the Qwen2.5-32B model significantly improves its inference capabilities through reinforcement learning. Core metrics such as mathematical code (AIME 24/25, livecodebench) and some general metrics (IFEval, LiveBench, etc.) reach the full-fledged level of DeepSeek-R1, and all metrics significantly surpass those of DeepSeek-R1-Distill-Qwen-32B, which is also based on Qwen2.5-32B.

Qwen3.5系列397B-A17B原生视觉语言模型，基于混合架构设计，融合了线性注意力机制与稀疏混合专家模型，实现了更高的推理效率。在语言理解、逻辑推理、代码生成、智能体任务、图像理解、视频理解、图形用户界面（GUI）等多种任务中，均展现出与当前顶尖前沿模型相媲美的卓越性能。具备强大的代码生成与智能体能力，对于各类智能体场景具有良好的泛化性。

Qwen3.5原生视觉语言系列Flash模型，基于混合架构设计，融合了线性注意力机制与稀疏混合专家模型，实现了更高的推理效率。模型效果在纯文本与多模态方面相较3系列均实现飞跃式进步；响应速度快，兼具推理速度和性能。

The Qwen3.5 native vision-language Plus models adopt a hybrid architecture that combines linear attention mechanisms with sparse mixture-of-experts models, achieving higher inference efficiency. Across a wide range of benchmark tasks, the 3.5 series demonstrates outstanding performance comparable to today’s leading frontier models, while delivering a significant leap over the 3 series in both pure text and multimodal capabilities.

The latest Qwen3-Max-Preview model is the preview version of the Max model in the Qwen3 series. Compared with the Qwen 2.5 series, it delivers substantial overall improvements in general capabilities, with significantly enhanced Chinese and English text understanding, complex instruction following, subjective open-ended task performance, multilingual ability, and tool-use capability. The model also produces fewer knowledge hallucinations.

Step 3.5 Flash is the strongest open-source foundation model released by StepFun to date. Its capabilities in agent scenarios and mathematical tasks approach those of closed-source models, enabling it to handle complex, long-chain tasks effectively.

OpenAI's third-generation embedding model, more powerful than the small version, suitable for tasks requiring the highest performance

Claude Sonnet 5 is built as the most dominant sonnet model to date. It can devise plans, utilize tools such as browsers and terminals, and operate autonomously, achieving a level that was previously only attainable by larger and more expensive models just a few months ago. For many developers, the era of agent-based AI began with Sonnet-like models: Claude Sonnet 3.5, 3.6, and 3.7 were the first models to demonstrate impressive coding and tool usage capabilities. However, recently, the most significant advancements in agent capabilities have emerged in our Opus-level models.

You can start with GPT-5.6 Sol for complex reasoning and coding, choose GPT-5.6 Terra to balance intelligence and cost, or use GPT-5.6 Luna to handle cost-sensitive, high-traffic tasks.

You can start with GPT-5.6 Sol for complex reasoning and coding, choose GPT-5.6 Terra to balance intelligence and cost, or use GPT-5.6 Luna to handle cost-sensitive, high-traffic tasks.

You can start with GPT-5.6 Sol for complex reasoning and coding, choose GPT-5.6 Terra to balance intelligence and cost, or use GPT-5.6 Luna to handle cost-sensitive, high-traffic tasks.

Kimi K2.7 Code is Kimi’s most intelligent Coding model, capable of completing programming tasks with higher success rates in long context. It features a native multimodal architecture that supports text, image, video input, and thinking modes, and dialogue and agent tasks. Kimi K2.7 Code HighSpeed is the high-speed version of Kimi K2.7 Code, the same model as Kimi K2.7 Code, but with an output speed of approximately 180 Tokens/s and up to 260 Tokens/s in short context scenarios, delivering a more extreme coding experience. Context length 256k, supports long thinking and deep reasoning. Supports automatic context caching functionality, ToolCalls, JSON Mode, Partial Mode.

The Qianwen QVQ visual reasoning model supports visual input and thought chain output, demonstrating stronger abilities in mathematics, programming, visual analysis, creativity, and general tasks.

The Qianwen series code and programming model is a language model specifically designed for programming and code generation, with excellent performance and outstanding effects.

The Tongyi Qianwen 3 series Max model has undergone a special upgrade in the direction of intelligent agent programming and tool calling compared to the preview version. The official version of the model released this time has reached the level of SOTA in the field and is suitable for more complex intelligent agent requirements in scenarios.

The Qwen3.6 native visual language series Plus model demonstrates outstanding performance comparable to current cutting-edge models, with significantly improved model performance compared to the 3.5 series. The model has coding capabilities in Agent coding, front-end programming, Vibe coding, multimodal object recognition, and more OCR、 Significant enhancement in object positioning and other abilities.