点击卡片查看全部路由组价格，勾选 2～3 个模型即可加入对比。
GPT-6 Astra is the most capable model of OpenAI, built for the hardest end-to-end work. Use it for complex reasoning, coding, computer use, research, and document creation. reasoning.effort supports low, medium, high, xhigh, and max.
Gemini-3.8-Flash is the general availability (GA) flagship multimodal model, designed for code development and AI agent scenarios. It supports configurable reasoning depth, offers a million-token context window, and is compatible with image, text, audio, and video inputs. With standout capabilities in tool invocation and multi-step task execution, it is well-suited for large-scale production workloads.
GPT-5.6 Sol is the flagship high-end model of the GPT-5.6 product line, with specifications benchmarking the top-tier original GPT-5 without suffixes. You may call the API using the alias gpt-5.6-sol, and the platform will automatically route requests to the GPT-5.6 Sol service cluster. This model supports differentiated calls via multiple suffixes, which can be appended to select performance tiers: -high, -xhigh, -low, -medium, -max, -ultra.
Nano Banana Pro is the next-generation AI image generation and editing model in the Google Gemini series, an upgrade of Gemini 2.5 Flash Image (Nano Banana). The model integrates a multimodal Transformer and diffusion model, natively supports 2K (2048×2048) and 4K resolution output, and delivers significant improvements in image quality, text rendering, and physical reasoning.
GPT-5.5 is OpenAI's flagship large language model released on April 24, 2026, designed as a new intelligence for real-world work and agents. Its core breakthrough is autonomous planning and execution of multi-step complex tasks, excelling in coding, computer operations, and research analysis while achieving higher efficiency with fewer tokens. Appending high/low/medium/xhigh indicates the thinking effort level.
GPT Image 2 is our most advanced image generation model, supporting fast, high-quality image generation and editing. It supports flexible image dimensions and high-fidelity image input.
Gemini 3.1 is Google's most intelligent model family to date, grounded in advanced reasoning. It is designed to turn any idea into reality by mastering 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.
Claude Opus 5, the default model for Claude Max and the most powerful model for Claude Pro, is a thoughtful, proactive, cost-efficient agentic model that delivers nearly the cutting-edge intelligence of flagship Claude Fable 5 at half the cost and sets new state-of-the-art standards for coding and professional knowledge work.
Claude Sonnet 5 comes with a default 1 million-token context window (1 million tokens serves as both its default and maximum capacity; no variants with smaller context limits are available). It features a maximum output token limit of 128k, adaptive reasoning capabilities, and retains all the same tooling and platform functionalities as Claude Sonnet 4.6—with the sole exception of the Priority Tier feature, which is unsupported on Claude Sonnet 5.
claude-opus-4-8 is Anthropic's most capable model for sustained, complex long-horizon work. Ideal for developers tackling large projects, building Agents, or any scenario demanding high autonomy and output quality.
Claude Opus 4.6 is Anthropic's latest flagship AI model, delivering major advances in professional task handling, long-context understanding, and multi-agent collaboration.
Gemini 3.5 Flash is now GA with stable performance ready for production at scale. As our most intelligent Flash model, it delivers consistently leading performance at scale for agent execution, coding, and long-horizon tasks.
Gemini 2.5 Pro is Google's latest and most advanced AI model, excelling at coding and complex prompts. Through deep thinking, it reasons before responding to boost performance and accuracy. The model tops multiple benchmarks and ranks first on the LMArena leaderboard for reasoning and code generation. Supports multimodal input including text, images, audio, video, and code.
Nano Banana Pro is a reasoning-driven precision engine for professional-grade image editing and generation, delivering studio-level accuracy and advanced creative control. Best suited for complex graphic design, high-fidelity product visuals, and factual data visualization grounded in real-world information via Google Search — tasks that require accurate text rendering.
AI video generation model supporting text-to-video, image-to-video, and video continuation. Available in multiple versions including Video-O1 and V1.0 through V2.6, with standard (std) and high-quality (pro) modes. Supports durations of 5, 10, and 15 seconds. Some versions also support audio generation and voice customization.
Claude‑fable‑5 is a publicly available, high-performance large language model from Anthropic featuring ultra-long context, multimodal understanding, complex reasoning, and enterprise-grade knowledge work capabilities, backed by strict safety protections.
Nano Banana 2 delivers high-quality image generation and conversational editing at competitive prices with low latency. It is the efficient version of Gemini 3.1 Pro Image, optimized for speed and high-throughput developer use cases.
An update to Nano Banana 2 (Gemini 3.1 Flash Image), built for high-efficiency image generation and conversational editing with improved visual quality, multi-turn character consistency, accurate text rendering, and search-grounded generation across 1K, 2K, and 4K resolutions.
Gemini 3.5 Flash-Lite is a cost-efficient multimodal model that enables low-cost sub-agent task execution and document parsing. The model supports text, image, video, audio and PDF inputs, and is engineered for high-volume agent workflows, simple data extraction, and applications where latency and API costs constitute primary constraints.
GLM-5.3-Flash 同时进行了多项架构升级，是首个采用稀疏注意力与线性注意力混合架构的开源前沿模型，在保持精准长上下文能力的同时，大幅降低了长上下文服务成本。此外，GLM-5.3-Flash 还采用流形约束超连接（Manifold-Constrained Hyper-Connections，mHC），进一步提升了模型的 scaling 能力。
Gemini 3.6 Flash consistently delivers state-of-the-art model intelligence. Optimized to process real-world tasks at greater speed and lower cost, it is built for the agent era and excels at code generation, agent execution, and spatial reasoning. This model performs exceptionally well in fast agent loops involving complex coding cycles and iterations.
DeepSeek-V4-Flash is the lightweight variant of the DeepSeek V4 series, prioritizing cost-effectiveness and high throughput efficiency, suited for general conversation and basic text tasks, with support for 1M-token long context and efficient inference.
DouBao-Seedream-5-0-pro-260628 is ByteDance's latest image creation model, capable of precisely parsing complex instructions and visual content. It features enhanced world knowledge breadth, reference consistency, and professional-scene generation quality to better serve enterprise visual creation needs.
DeepSeek-V4-Pro is a high-performance open-source large model from DeepSeek with top-tier reasoning and Agent capabilities, ultra-long context support, compatibility with Ascend chips, and exceptional cost-effectiveness.
Kimi K3 是 Kimi 迄今能力最强的旗舰模型，拥有 2.8 万亿参数，基于 KDA 混合线性注意力机制（Kimi Delta Attention）和注意力残差（Attention Residuals）技术构建，原生支持视觉理解，并拥有 100 万 token 上下文窗口。它是全球首个开源的 3 万亿级别模型，面向长程编程、知识工作和推理等前沿智能场景而设计。
Suno API for building a platform equivalent to the official Suno website, supporting all official site operations. This model is for song generation and supports custom mode, inspiration mode, and song continuation.
Gemini 3 is Google's most intelligent model family to date, built on advanced reasoning capabilities. It is designed to turn any idea into reality by mastering Agent workflows, autonomous coding, and complex multimodal tasks. gemini-3-pro-preview is best suited for complex tasks requiring broad world knowledge and advanced cross-modal reasoning.
Google's latest advanced AI model. veo3.1 fast mode supports automatic audio generation alongside video. High quality at a low price — the best value option. Supports first/last frame control and text-to-video generation (8-second clips).
Kling O1 embraces the Multi-modal Visual Language (MVL) philosophy, using natural language as the semantic backbone combined with multimodal descriptions including video, images, and subjects to precisely understand your intent — making operation more intuitive and creation more efficient.
GLM-5.2 is the flagship model built for the long-task era. It supports a truly usable 1M context capable of holding project-level engineering context, delivers more stable long-horizon task execution and more reliable adherence to engineering standards, and can complete the full development pipeline from requirements to multi-platform deployment in a single task.
Claude Sonnet 4.5 is the world's best coding model, the most powerful model for building complex agents, and the top model for computer use. It shows significant advances in reasoning and math.
GLM-5.1 is a model designed for long-horizon tasks. It can work autonomously and continuously for up to 8 hours within a single task — independently planning, executing, and self-improving to deliver complete, engineering-grade outcomes.
Doubao-Seedream-5.0-lite is ByteDance's latest image creation model, capable of precisely parsing complex instructions and visual content. It features enhanced world knowledge breadth, reference consistency, and professional-scene generation quality to better serve enterprise visual creation needs.
GPT-4o is OpenAI's most advanced multimodal model, faster and cheaper than GPT-4 Turbo with stronger vision capabilities. The model features a 128K context window and a knowledge cutoff of October 2023. Models in the 1106 series and above support tool_calls and function_call.
The Qwen3.6 series 27B is a native vision-language Dense model. Compared to 3.5-27B, it focuses on improving agentic coding and further strengthens STEM and reasoning capabilities. On the vision side, it significantly enhances spatial intelligence, object localization and detection, with steady gains in video understanding, document OCR, and visual Agent capabilities.
Kimi K2.6 is Kimi's latest and most intelligent model, featuring stronger and more stable long-range code writing, significantly improved instruction following and self-correction, with support for text, image, and video input, thinking and non-thinking modes, and both dialogue and Agent tasks.
Our most cost-efficient multimodal model, delivering the fastest performance for high-frequency, lightweight tasks. Gemini 3.1 Flash-Lite is best suited for high-volume agentic workloads, simple data extraction, and ultra-low-latency applications where budget and speed are the primary constraints.
Google's latest advanced AI model. veo_3_1--components supports up to three reference image inputs for faster, more stable, and more controllable generation, redefining creative freedom in image-to-video. 4K model. (8-second duration)
Z-Image-Turbo is a high-efficiency image generation model ranked #1 among open-source text-to-image models on Artificial Analysis benchmarks. With just 6 billion parameters and 8-step inference, it produces photorealistic images on par with large-scale commercial models, excelling in bilingual (Chinese and English) text rendering, complex semantic understanding, and diverse subject generation.
Claude Opus 4.5 is our most intelligent model to date. It sets new benchmarks in coding, agentic tasks, computer use, and enterprise workflows. Opus 4.5 represents a meaningful advance in AI system capabilities.
Google's latest advanced AI model. veo3.1 high-quality mode supports automatic audio generation alongside video. High quality at a low price — the best value option. Supports first/last frame control and text-to-video generation (8-second clips).
Seedream 4.5 natively supports text, single-image, and multi-image inputs, enabling subject-consistent multi-image fusion creation, image editing, and image sequence generation — making image creation freer and more controllable.
A cost-effective Plus model in the Qwen3.7 series with strong text capabilities, fully upgraded vision-language understanding, and complete agentic capabilities for coding, tool use, and productivity workflows. Its core strength is multimodal interactive hybrid Agent capability — perceiving real-world scenes, reading screens and operating GUIs, generating code from visual references, and navigating mobile apps end-to-end.
Qwen VL is a text-generation model with visual (image) understanding. Beyond OCR, it can summarize and reason over images — for example, extracting product attributes from photos or solving problems based on diagrams.
A digital avatar video generation model that converts static portrait images into dynamic talking videos. Supports standard (std) and high-quality (pro) modes, billed per second. Suitable for virtual presenters, video dubbing, and digital avatar production.
MiMo-V2.5-Pro is designed for complex task scenarios with deep optimization for Agent and Coding applications, ranking #1 among global open-source models on the GDPVal-AA and ClawEval leaderboards.
kimi-k2 is a MoE foundation model with exceptional coding and Agent capabilities. In benchmarks across key categories — programming, math, and Agent tasks — K2 outperforms other leading open-source models.
GPT-5-Codex is a version of GPT-5 optimized for agentic coding in Codex. Like its predecessor codex-1, the model is trained on real-world coding tasks using reinforcement learning across diverse environments to produce code that closely matches human style and PR preferences, follow instructions precisely, and iteratively run tests until they pass.
Doubao-Seed-2.0-lite is a balanced model for high-frequency enterprise scenarios, trading performance against cost. It surpasses the previous-generation Doubao-Seed-1.8 in overall capability, handling unstructured information processing, content creation, search and recommendation, and data analysis. Supports long context, multi-source information fusion, multi-step instruction execution, and high-fidelity structured output — delivering consistent results at significantly reduced cost.
Qwen3.5 series 35B-A3B native vision-language model built on a hybrid architecture combining linear attention with sparse MoE, achieving higher inference efficiency. Its overall performance is close to Qwen3.5-27B.
gpt-4o-mini-2024-07-18. GPT-4o Mini is designed to be affordable and capable. It offers a 128K context window, up to 16K output tokens per request, and a knowledge cutoff of October 2023. Models in the 1106 series and above support tool_calls and function_call.
Kling 2.6 Motion Control AI motion transfer technology — real movements, precise control. Transfers exact motion from a reference video onto a static character image. Generates up to 30 seconds of animation in a single pass, with full-body motion accuracy, precise gesture control, and consistent character identity.
GPT-5.5pro is available for Responses API requests, including via the Batch API, supporting multi-turn model interactions before responding, with additional advanced API features coming in the future. As GPT-5.5pro is designed for complex problems, some requests may take several minutes to process. To avoid timeouts, use background mode.
The largest and most capable Max model in the Qwen 3.7 series, currently offering pure-text model capabilities. Qwen 3.7 is a next-generation flagship model built for the agentic era, with core strengths in the breadth and depth of agent capabilities: excelling at coding, office productivity, and long-horizon autonomous task execution.
kimi-k2.5 is Kimi's most capable model to date, built on a native multimodal architecture with support for both vision and text input, thinking and non-thinking modes, and both conversational and Agent tasks.
According to Anthropic's announcement, the new model outperforms OpenAI o3 across agentic tool use, agentic coding, and multilingual Q&A. Opus 4.1 achieves a top coding score of 74.5% on SWE-bench Verified. This release also enhances Claude's deep research and data analysis capabilities, particularly in detail tracking and agentic search.
GPT-4O is OpenAI's latest language model with enhanced comprehension and generation capabilities. It excels across multiple NLP domains, including dialogue generation, text summarization, and information retrieval. GPT-4O integrates more advanced techniques to understand complex contexts and produce more coherent, relevant responses. The model also deepens its grasp of specialized domain content, supports multiple languages and styles, and improves the overall user experience. Its goal is to deliver more accurate and practical language interactions.
内容丰实：支持最大 4.5k token 输入，支持图中图密集信息排版，让报纸、分镜、菜单、试卷等复杂版面一次生成。 细节真实：支持 10px 小字精准渲染，微表情、毛孔、发丝等细节生动还原，逼近真实摄影的质感。 知识厚实：支持 12 国语言、多种字体原生渲染，主流网页、游戏、直播等界面仿真，外部知识全纳入。 Qwen-Image-3.0-Pro 不只是在追求"好看"，更在追求“好用”——让图像生成真正成为可落地的生产力工具。
The Lipsync API is designed to solve lip-sync issues in video. It analyzes the mouth movements of speakers and aligns them precisely with provided audio or TTS output, making videos more expressive and engaging while deepening storytelling impact.
The Qwen3.6 native vision-language Plus model delivers performance rivaling today's top frontier models, with a significant quality leap over the 3.5 series. Notable improvements include agentic coding, frontend programming, vibe coding, multimodal universal recognition, OCR, and object localization.
Doubao-Seed-2.0-pro is a flagship all-purpose general model designed for the Agent era, tackling complex reasoning and long-horizon task execution. It emphasizes multimodal understanding, long-context reasoning, structured generation, and tool-augmented execution. Excels at complex instruction following and multi-constraint tasks, reliably handling multi-step planning, complex vision-language reasoning, video content understanding, and demanding analytical workloads.
GPT-5.4-pro uses more compute to think more deeply and consistently deliver better answers. Accessible only through the Responses API to support multi-turn model interactions prior to responding to API requests, as well as other advanced API features in the future.
Grok 4.20 is xAI's latest flagship model, featuring industry-leading speed and agentic tool-calling capabilities. It combines the lowest hallucination rate on the market with strict prompt compliance, consistently delivering accurate and truthful responses.
混元Hy4 preview总参数770B、激活参数49B，重点面向Agent、Coding与生产力场景优化，强化复杂任务中的理解、规划、工具调用与持续执行能力。相比上一代模型，Hy4 preview在多步骤Agent、代码开发和生产力任务上进一步提升，具备更好的任务拆解、上下文承接、指令遵循和长链路执行能力。在Coding场景中，进一步增强代码理解、生成、修改和复杂工程任务处理能力；在生产力场景中，重点提升文档处理、信息分析、办公自动化、游戏开发、网页生成及跨工具协作等能力。Hy4 preview适用于Coding Agent、复杂工具调用以及各类需要多步骤规划和持续执行的Agent工作流，为真实业务场景中的复杂任务提供更可靠的任务完成能力。
Grok 4.20 is xAI's latest flagship model, featuring industry-leading speed and agentic tool-calling capabilities. It combines the lowest hallucination rate on the market with strict prompt compliance, consistently delivering accurate and truthful responses.
o3-mini is OpenAI's latest small reasoning model, optimized for coding, math, and science tasks. o3-mini responds faster than o1-mini and can match or exceed o1 accuracy at higher reasoning settings.
Qwen3.5 series 122B-A10B native vision-language model built on a hybrid architecture combining linear attention with sparse MoE for higher inference efficiency. Its overall performance is second only to Qwen3.5-397B-A17B; text capabilities significantly surpass Qwen3-235B-2507, and vision capabilities surpass Qwen3-VL-235B.
Kimi K2.7 Code is Kimi's most capable coding model to date. It follows instructions more reliably in long contexts and completes programming tasks at a higher success rate. Supports text, image, and video input; thinking mode only; suitable for conversation and Agent tasks.
Grok 4 features deep reasoning capabilities, trained on xAI's Colossus supercomputer, delivering stronger logical reasoning and text generation. xAI claims it is the world's most powerful AI model, reaching PhD-level performance on academic tasks. It excels in real-time speed, reasoning, and advanced vision.
gpt-4o-2024-08-06 features a 128K context window, up to 16K tokens output per request, and a knowledge cutoff of October 2023. It supports stricter JSON-formatted output and stronger reasoning. Models in the 1106 series and above support tool_calls and function_call.
Seedream 4.0 is a multimodal image generation model from ByteDance. Within a unified framework, it integrates text-to-image, image editing, and multi-image composition: it accepts up to 10 images per input for composite editing, and through deep reasoning over prompts automatically selects the optimal aspect ratio and generation count — outputting up to 15 thematically linked images in a single run. The model also significantly improves Chinese text generation accuracy and content diversity, supports 4K ultra-HD output, and delivers an end-to-end solution from generation to editing for professional image creation.
The MiniMax speech large model intelligently predicts emotion and intonation from context to generate ultra-natural, high-fidelity, personalized voice. It excels across social, podcast, audiobook, news, education, and digital human scenarios.
Seedance 1.0 pro fast is a comprehensive model that achieves an outstanding balance of video generation quality, speed, and price. It inherits the core strengths of Seedance 1.0 pro while delivering faster generation speeds and more competitive pricing, giving creators a dual optimization in efficiency and cost.
Qwen massively scaled language model supporting Chinese, English, and other languages. Suited for text creation, text processing, coding assistance, translation, and dialogue simulation.
Doubao-Seed-2.1 是面向 Coding 和 Agent 时代打造的新一代大模型，提供 Pro 和 Turbo 两个版本，分别面向高复杂度任务探索和规模化生产场景，Seed 2.1 在 Coding 工程交付、Agent 长链路任务执行和多模态理解三大方向上全面升级，以更强的自主规划与动态修复能力，胜任真实研发与高价值生产任务。
The latest model gemini-embedding-2-preview is the first multimodal embedding model in the Gemini API. It maps text, images, video, audio, and documents into a unified embedding space, enabling cross-modal search, classification, and clustering in over 100 languages.
HappyHorse-1.0-R2V supports reference-to-video generation with more stable subject and scene referencing, up to 9 reference images, precise preservation of creative intent, and stronger expressive capabilities.
Qwen3.5 native vision-language Plus model built on a hybrid architecture combining linear attention with sparse MoE for higher inference efficiency. Across multiple benchmarks, the 3.5 series matches today's top frontier models, with significant capability leaps over the 3 series in both text and multimodal tasks. This is a snapshot from February 15, 2026.
A frontier AI video foundation model with core capabilities including Text-to-Video and Image-to-Video. Supports uploading images or entering prompts to instantly generate high-definition dynamic video. Widely used for short-form video, advertising, and content creation.
Qwen3.5 native vision-language Plus model built on a hybrid architecture combining linear attention with sparse MoE for higher inference efficiency. Across multiple benchmarks, the 3.5 series matches today's top frontier models, with significant capability leaps over the 3 series in both text and multimodal tasks.
GPT-5.4 Pro uses more compute to think deeper and consistently deliver better answers. Accessible only through the Responses API to support multi-turn model interaction before fulfilling Responses API requests, as well as other advanced API features in the future.
Seedream 3.0 is a bilingual Chinese and English image generation foundation model with native high-resolution support. Its overall capability is on par with GPT-4o, placing it in the global top tier. Features faster response, more accurate small-text generation with improved text layout, strong instruction following, enhanced aesthetics and structure, and high fidelity with detailed rendering.
o3-mini is OpenAI's latest small reasoning model, optimized for coding, math, and science tasks. It responds faster than o1-mini and can match or exceed o1 accuracy at higher reasoning settings.
HappyHorse-1.0-Video-Edit supports video editing via natural-language instructions, with up to 5 reference images for local or global editing of video elements, accurate replication of video motion, and stronger expressive capabilities.
HappyHorse-1.0-T2V supports text-to-video generation with highly faithful dynamic scene rendering, precise understanding of text semantics, and smooth, natural, detail-rich, high-quality video output.
Gemini-TTS is the latest advancement in our Cloud TTS technology. Beyond generating natural speech, it enables fine-grained control over the generated audio through text-based prompts. With Gemini-TTS you can synthesize single-speaker or multi-speaker voices — from short clips to long-form narration — and precisely specify style, accent, speaking rate, intonation, and emotional expression, all controlled flexibly through natural language prompts.
The first omnimodal understanding model in the Doubao family, supporting native unified understanding of video, images, audio, and text, with upgraded Agent, Coding, and GUI capabilities.
The Wan 2.7 flagship image generation and editing model, supporting text-to-image, text-to-image-set, image-to-image-set, image editing, multi-reference generation, and interactive editing, with stronger performance in text rendering, subject consistency, and complex instruction following.
An AI audio generation model supporting text-to-sound-effects and video dubbing. Generates sound effects from text descriptions or automatically dubs videos. Suitable for sound design and video post-production.
The Gemini API provides text embedding models that generate embeddings for words, phrases, sentences, and code. Embedding tasks such as semantic search, classification, and clustering deliver more accurate, context-aware results than keyword-based approaches.
Qwen3 series Flash model that effectively blends thinking and non-thinking modes, switchable mid-conversation. Excels at complex reasoning tasks with significantly improved instruction following and text comprehension. Supports 1M context length with tiered pricing based on context length.
doubao-seed-1.8 delivers stronger multimodal understanding and Agent capabilities, achieving outstanding performance across many complex real-world tasks and further helping enterprises create value.
HappyHorse-1.0-I2V supports image-to-video generation with highly faithful dynamic scene rendering, precise text semantic understanding, and high-quality output that is smooth, natural, and detail-rich.
An accelerated model in the Qwen-Image-2.0 series that unifies image generation and image editing. Features professional text rendering with up to 1k token instruction support, finer realism for detailed photorealistic scenes, and stronger semantic adherence. The accelerated variant achieves the optimal balance between model quality and performance.
Public cloud version. Qwen's best-performing model, designed for complex, multi-step tasks. A hundred-billion-parameter large language model supporting Chinese, English, and other languages. Ideal for text creation, text processing, coding assistance, translation, and conversational simulation.
Seedance 1.0 is the latest video generation foundation model from the Doubao large model team at ByteDance. Seedance 1.0 Pro, the high-parameter variant of the series, features unique multi-shot narrative capability and excels across multiple dimensions. It achieves breakthroughs in semantic understanding and instruction following, producing 1080P HD videos with fluid motion, rich detail, diverse styles, and cinematic aesthetics.
The Qwen3.6 series 35B-A3B is a native vision-language model built on a hybrid architecture combining linear attention and sparse MoE, achieving higher inference efficiency. Compared to 3.5-35B-A3B, it delivers significant improvements in agentic coding, math reasoning, code reasoning, spatial intelligence, object localization, and object detection.
o1 full version. It spends more time thinking (forming a chain of thought) before arriving at an answer, making it better suited for complex reasoning tasks, especially in science and mathematics.
This is the December 1, 2025 snapshot. Compared to the July 28 snapshot, reasoning capability is improved; Agent capabilities and multi-turn tool calling are further enhanced; performance on subjective creative tasks is better. Supports 1M context length with tiered pricing based on context length.
Qwen3.5 397B-A17B native vision-language model built on a hybrid architecture combining linear attention with sparse MoE for higher inference efficiency. It delivers performance comparable to today's top frontier models across language understanding, logical reasoning, code generation, Agent tasks, image understanding, video understanding, and GUI tasks, with strong generalization across diverse Agent scenarios.
The Modify feature lets you edit any portion of an existing video — adding, replacing, removing, or modifying elements, or changing the overall style. It helps you quickly produce derivative works, turning one video into many new creative possibilities.
GLM-4.5 and GLM-4.5-Air are the latest flagship model series, built as foundation models for Agent applications. Both use a Mixture-of-Experts (MoE) architecture. GLM-4.5 has 355B total parameters with 32B active; GLM-4.5-Air uses a leaner design with 106B total parameters and 12B active.
An advanced lip-sync model with face detection and lip-sync capabilities. Delivers high-precision lip synchronization for natural, realistic speech. Face detection is billed per call; lip sync is billed in 5-second units.
GPT-5.1-chat-latest is a general-purpose model optimized for smooth conversation and assisted content creation in natural-language interaction and instant-messaging scenarios.
GPT-4 Turbo with Vision is the latest-generation model. More capable, with a knowledge cutoff updated to April 2023 and a 128k context window. Accepts text or image input and outputs text, solving difficult problems with greater accuracy than any previous model. Models in the 1106 series and above support tool_calls and function_call.
A text reranking model trained on the Qwen LLM foundation. It ranks the relevance of an input query against candidate documents, supports 100+ languages and long-text input, and is suited for text retrieval and RAG scenarios. Performance is aligned with the open-source Qwen3-Rerank model series.
The Multi-transition feature lets you supply 2–7 keyframes to generate a 1–30 second video with consistent style and smooth scene transitions. Designed for professional creators, it provides finer control over video sequences, keeping characters and actions coherent and controllable throughout.
GPT-4 Turbo with Vision is the latest-generation model. More capable than its predecessors, it features an April 2023 knowledge cutoff, a 128k context window, and accepts text or image inputs, solving difficult problems with greater accuracy.
Meta-Llama-3-70B-Instruct is the 70B instruction-tuned version, suited for conversational use cases, with superior performance in understanding linguistic nuance, context, and complex task execution.
Gemini Pro is Google AI's latest model, with updated guidelines released on September 26, 2025. Described as state-of-the-art, it offers strong reasoning and coding capabilities for complex tasks, with controllable thinking.
BAAI/bge-reranker-v2-m3 is a lightweight multilingual reranking model built on bge-m3. It delivers strong multilingual capabilities, easy deployment, and fast inference. The model takes a query and document as input and outputs a similarity score directly rather than embedding vectors. It suits multilingual scenarios, with particularly strong performance in Chinese and English.
The Restyle feature instantly transforms your video clips into a completely new visual style — such as 3D, anime, cinematic, or enhanced realism. Whether you're aiming for an artistic look or a branded visual identity, Restyle gives you the power to create your own world.
A role-play model in the Qwen series. This dynamically updated version is optimized for anthropomorphic role-play, with enhanced capabilities for following persona-specific instructions, advancing conversation topics, and empathetic listening. Supports deep reconstruction of personalized characters.
The MiniMax speech large model intelligently predicts emotion and intonation from context to generate ultra-natural, high-fidelity, personalized voice. It excels across social, podcast, audiobook, news, education, and digital human scenarios.
Open-source QwQ. A QwQ reasoning model trained on Qwen2.5-32B with substantially enhanced reasoning via reinforcement learning. Core benchmarks in math and coding (AIME 24/25, LiveCodeBench) and general benchmarks (IFEval, LiveBench, etc.) match DeepSeek-R1 at full capacity, surpassing DeepSeek-R1-Distill-Qwen-32B across all metrics.
A custom voice creation model that lets users create personalized voice profiles by uploading audio files or using audio from existing videos. The created custom voices can be used for subsequent text-to-speech synthesis, dubbing, and other applications.
The Qwen3.6 native vision-language Plus model delivers performance rivaling today's top frontier models, with a significant quality leap over the 3.5 series. Notable improvements include agentic coding, frontend programming, vibe coding, multimodal universal recognition, OCR, and object localization.
The sound effect generation (background music) API intelligently generates sound effects and ambient audio matched to video footage, and also accepts text-based instructions. Supports multiple scenarios and accurately matches video pacing and content logic.
The MiniMax speech large model intelligently predicts emotion and intonation from context to generate ultra-natural, high-fidelity, personalized voice. It excels across social, podcast, audiobook, news, education, and digital human scenarios.
The Qwen image generation Max series delivers excellent results across a wide range of generation tasks. Compared to the Plus series, it significantly reduces the AI-generated look of images, enhancing realism with more lifelike human textures, finer natural details, and improved text rendering.
The Qwen image generation Max series delivers excellent results across a wide range of generation tasks. Compared to the Plus series, it significantly reduces the AI-generated look of images, enhancing realism with more lifelike human textures, finer natural details, and improved text rendering.
Doubao-Seed-2.0-Code is optimized for enterprise-level programming needs. Building on Seed 2.0's strong Agent and VLM capabilities, it significantly enhances code generation — excelling at frontend development while offering specialized optimization for the multi-language coding scenarios common in enterprise environments. Ideal for integration with AI programming tools.
SparkDesk-v1.1 is a next-generation cognitive AI large model from iFlytek Spark, with cross-domain knowledge and language understanding. It understands and executes tasks through natural conversation, offering language comprehension, knowledge Q&A, logical reasoning, math problem solving, and code understanding.
Qwen3-Reranker-0.6B is a text reranking model from the Qwen3 series. It is designed to optimize the results of initial retrieval systems, with its core function being to re-rank documents based on their relevance to a given query. The model has 0.6B parameters and a 32k context length, and inherits the strong multilingual (100+ languages), long-text understanding and reasoning capabilities of its Qwen3 base model. Evaluation results show that Qwen3-Reranker-0.6B achieves excellent performance across multiple text retrieval benchmarks including MTEB-R, CMTEB-R, and MLDR.
GLM-4.5 and GLM-4.5-Air are the latest flagship model series, purpose-built as foundation models for Agent applications. Both use a Mixture-of-Experts (MoE) architecture. GLM-4.5 has 355B total parameters with 32B active; GLM-4.5-Air uses a more compact design with 106B total parameters and 12B active.
GPT-5.1-codex is the well-equipped heavy-duty powerhouse, designed to tackle the toughest coding challenges, handle large-scale codebases, and provide deep architectural guidance — the go-to model for enterprise-grade and highly complex development scenarios.
SWAP is a powerful video editing feature. Upload or generate a video, select a frame, and choose the object you want to edit. Then upload a target image and the system generates a brand-new video. Use it together with the mask/selection API to specify the region to replace. External mask_url input is not currently supported.
Mimic transfers motion from a reference video onto a target character image to reproduce and reconstruct actions. Provide a dance (or action) reference video and a character image; Mimic analyzes the motion sequence and rebuilds it on the supplied character, generating a new video in which the character reproduces the original movements.
Pixverse Mask Selection (also known as Swap Mask Generation) is an AI-powered smart mask generation tool in the PixVerse platform's Modify feature. It precisely selects regions to edit in a video — subjects, objects, backgrounds, and more — and applies targeted modifications using prompts, reference images, or replacement modes.
GPT-5.2 pro is available exclusively in the Responses API to support multi-turn model interactions and advanced API features. Because GPT-5.2 pro is designed to tackle hard problems, some requests may take several minutes to complete.
The MiniMax speech large model intelligently predicts emotion and intonation from context to generate ultra-natural, high-fidelity, personalized voice. It excels across social, podcast, audiobook, news, education, and digital human scenarios.
Doubao-Seed-2.0-mini targets low-latency, high-concurrency, and cost-sensitive scenarios, delivering maximum inference speed. Performance is on par with Doubao-Seed-1.6. Supports 256K context, four thinking-depth levels, and multimodal understanding — ideal for lightweight tasks where speed and cost are the priority.
The Claude 4 series is Anthropic's latest generation of AI models, representing the current pinnacle of large language model technology. This generation spans two product lines—Claude Sonnet 4 and Claude Opus 4—targeting different use cases and requirements.
A versatile all-around model and the first in China to support 256K context with thinking capability. Features deep thinking, multimodal understanding, and GUI operation. Users can flexibly enable or disable deep thinking as needed; adaptive thinking mode automatically determines whether to activate deep thinking based on prompt complexity, improving output quality while significantly reducing token consumption.
Supports adjustable thinking depth via the reasoning_effort field in the API, with four modes: minimal, low, medium, and high. This compresses overall output length to balance quality, latency, and cost across different use cases.
Ultra-fast response variant in the Doubao 1.6 series. Features deep thinking and multimodal understanding with 256K context support. Extremely low latency — TOPT (Top-of-Pipeline Time) as low as 10 ms — for near-instant responses. Visual understanding performance is on par with flagship models from other leading vendors. Ideal for latency-sensitive scenarios such as real-time interaction and vision task processing.
The enhanced deep-thinking variant of the Doubao 1.6 series, focused on deep reasoning. Further improves coding, math, and logical reasoning to handle more complex tasks. Supports 256K context for understanding and generating longer content, ideal for in-depth analysis and complex reasoning scenarios.
Significantly enhanced thinking capability. Compared to the Doubao 1.5-generation deep-understanding model, it further improves core skills in coding, math, and logical reasoning, and adds visual understanding support.
Doubao-Seed-1.6-vision is a vision deep-thinking model with stronger general multimodal understanding and reasoning for education, image review, inspection and security, and AI search Q&A scenarios. Supports a 256K context window with a maximum output length of 64K tokens.
SparkDesk-v2.1 is a next-generation cognitive AI large model from iFlytek Spark. It features cross-domain knowledge and language understanding, enabling task execution through natural conversation, with capabilities in language understanding, knowledge Q&A, logical reasoning, math problem solving, and code understanding.
A next-generation cognitive AI large model from iFlytek Spark, with cross-domain knowledge and language understanding. It comprehends and executes tasks through natural conversation, offering language understanding, knowledge Q&A, logical reasoning, math problem solving, and code comprehension.
Next-generation cognitive AI large model from iFlytek Spark, with cross-domain knowledge and language understanding. Executes tasks through natural conversation, covering language understanding, knowledge Q&A, logical reasoning, math problem solving, and code comprehension.
可灵图片 3.0 是快手自研的高清 AI 图像生成模型，支持文生图、图生图、多图参考、风格转绘、人像 / 角色参考、局部重绘。具备影视级叙事画面表达能力，精准响应创作指令，支持 2K/4K 超高清图像直出，画面细节、光影、材质一致性强，适合影视分镜、概念原画、电商海报、商业插画等图像创作场景，接口采用异步任务调用模式。生成图像时请指定参数"resolution"为大写的"2K"或"4K", 模型名使用 "kling"
The Audio API lets developers build speech-to-text and text-to-speech features. It supports speech recognition and generation in multiple languages, suitable for building voice assistants, transcription apps, and more.
Kling AI effects hub offering 136+ creative video effects, including holiday themes, dynamic effects, and style transfers. Pricing varies by scene, ranging from 1 to 7 RMB per use. Some effects support double-effect stacking.
A multi-element compositing model that intelligently blends multiple elements within a video for complex creative productions. Designed for advanced video creation requiring multi-layered, multi-element compositions.
Kling-omni-image is a complete AI creative suite for video and image creation. Kling Omni integrates video and image generation into a single unified creative suite, designed for professionals who need consistently high-quality output every time.
bce-reranker-base_v1 is a bilingual and cross-lingual reranking model developed by NetEase Youdao, supporting Chinese, English, Japanese, and Korean. Used in RAG systems to precisely rerank retrieval results, it provides meaningful relevance scores to help filter low-quality passages. Optimized for a range of RAG tasks including translation, summarization, and Q&A, the model requires no task-specific instructions, offers broad domain adaptability, and has been validated across multiple Youdao products.
o3-pro is a powerful reasoning model from OpenAI and an upgrade to o3. It excels at handling complex problems and delivering precise answers, with clear advantages in scientific research, coding, education, and writing.
Qwen3's training dataset is significantly expanded compared to Qwen2.5. While Qwen2.5 was pre-trained on 18 trillion tokens, Qwen3 uses nearly twice as much data—approximately 36 trillion tokens—covering 119 languages and dialects.
模型名称会变化，选型方法可以长期复用：先明确质量下限，再比较真实成本，最后用智能路由把简单任务分给轻量模型。
适合代码生成、仓库理解、工具调用和需要连续执行多个步骤的自动化任务。
适合改写、摘要、营销文案、客服回复和长篇中文内容处理。
适合多步骤分析、方案规划、数学逻辑和对答案可靠性要求较高的任务。
适合分类、抽取、标签生成、翻译和大量规则清晰的重复请求。
适合读取截图、图表、商品图片和带有视觉信息的文档页面。
适合营销素材、概念图、商品视觉和短视频内容的快速生产。
