apiex选择模型后可复制配置给 AI，也可以快速测试当前连接延迟。
Anthropic
Claude Fable 5 is Anthropic’s advanced long-horizon reasoning model, supporting text, image, and file inputs with text output. It features reasoning capabilities and a 1M-token context window, making it suitable for autonomous knowledge work, software development, and complex multi-step tasks.
Anthropic
Fast and affordable. Handles simple tasks well.
Anthropic
Strong in deep reasoning and long-context tasks.
Anthropic
Deepest reasoning in the Claude family. 1M context.
Anthropic
Deepest reasoning in the Claude family. 1M context.
Anthropic
Built for long-running async agents, excelling at complex multi-step tasks and end-to-end workflow execution.
Anthropic
Built for long-running async agents, excelling at complex multi-step tasks and end-to-end workflow execution.
Anthropic
Opus family flagship with reasoning. Suited for long-horizon autonomous agents, large-codebase analysis, and multimodal knowledge work.
Anthropic
Opus family flagship with reasoning. Suited for long-horizon autonomous agents, large-codebase analysis, and multimodal knowledge work.
Anthropic
Flagship model for demanding reasoning, coding, and long-horizon agentic work, spanning end-to-end software, code review, parallel subagent coordination, visual analysis of charts and documents, and complex office deliverables.
Anthropic
Previous-gen Sonnet. Still strong for long-form content.
Anthropic
Best balance of quality and cost from Anthropic.
Anthropic
Coding and agent quality approach the frontier, with native reasoning built in, well-suited for complex planning and long-horizon professional tasks that require cross-step consistency.
baidu
A natively omni-modal model with unified understanding of text, image, audio, and video, providing a flagship foundation for full-modality capabilities.
baidu
An efficient iteration of the flagship line, delivering major gains in reasoning, agentic handling, and search with far fewer parameters, balancing high performance and low-cost deployment.
bytedance
Turbo tier in Seed 2.1: multimodal model for coding and long-horizon agents, suited to multi-step task execution and visual content understanding.
bytedance
Character roleplay model for virtual companion scenarios, with stable persona adherence and plot progression in multi-turn dialogue, supporting multimodal inputs.
bytedance
Agent-oriented foundation model tuned for tool calling and complex instruction following, with use cases spanning GUI agents, search agents, and multi-step task orchestration.
bytedance
Code-focused preview model for agent workflows, with competitive SWE-Bench and LiveCodeBench scores, aimed at autonomous coding agents and multi-language software tasks.
bytedance
Lightweight tier of the Seed 2.0 family tuned for low-latency agent, coding, and GUI workloads.
bytedance
Lightweight tier in the Seed 2.0 family, tuned for high-throughput, low-latency tasks with adjustable reasoning levels. Fits batch processing, moderation, and classification.
bytedance
Flagship of the Seed 2.0 series for complex reasoning and long-horizon agent workflows with reliable tool use.
DeepSeek
Open-source reasoning model with visible chain-of-thought, tuned for math, coding, and multi-step problem solving.
DeepSeek
Open-source general-purpose LLM with representative instruction-following and coding skills within the open-weights ecosystem, suited for chat, code assistance, and enterprise text workflows.
DeepSeek
Hybrid model that switches between thinking and non-thinking modes on a single endpoint, post-trained for tool use and code agent workflows.
DeepSeek
Best value for code and math. Fraction of flagship cost.
DeepSeek
Experimental interim release focused on long-context efficiency, with optional reasoning mode and V3.1-level performance overall.
DeepSeek
Faster, lower-cost variant of DeepSeek V4, optimized for coding assistants, high-throughput chat, and latency-sensitive Agent workflows.
DeepSeek
DeepSeek V4 flagship with adjustable reasoning depth, built for full-codebase analysis, multi-step automation, and complex reasoning tasks.
A high-performance general-purpose model from Google, designed for advanced reasoning, coding, mathematics, and scientific tasks. Its built-in thinking capabilities improve response accuracy and enable deeper contextual understanding.
A lightweight reasoning model in the Gemini 2.5 family, optimized for ultra-low latency and cost efficiency, with faster token generation and higher throughput. Thinking is disabled by default but can be enabled via API.
Massive context window. Great for video understanding.
Speed-tier variant in the Gemini 3 line, pairing Pro-grade reasoning with Flash-level latency and cost, built for agent workflows and high-throughput interactive apps.
A high-efficiency multimodal lite model for low-latency, high-volume workloads like translation, classification, and data extraction, priced at about half of Gemini 3 Flash.
Google's latest. Strong reasoning with native image and video.
Designed for efficient multimodal AI tasks, offering strong coding, reasoning, real-time chat, and agent execution at Flash-tier cost and speed.
Fast and cost-efficient.
Open-weights dense model with stronger math and coding, plus native function calling for self-hosted agent workflows.
Built for fast generation and conversational editing; low latency and cost; multimodal text-and-image input.
Mid-tier image generation and editing model in the series, with near-flagship visual quality at higher generation speed. Renders legible text and photorealistic subjects, suited for infographics and marketing visuals.
The lightest image model in its series, producing images in about four seconds with consistent character rendering and precise edits. Suited for high-throughput pipelines and interactive visual iteration.
Built-in reasoning; complex multi-turn creation; up to 4K; optional web grounding.
liquid
Instruction-tuned optimization for precise execution.
minimax
Cost-effective general model for balanced speed and quality.
minimax
Built for conversation and bilingual chat.
minimax
A high-throughput variant of M2.5 with matching quality at roughly triple the speed, tuned for coding and agentic tool use under latency-sensitive workloads.
minimax
A next-generation autonomous language model that uses multi-agent collaboration to plan, execute, and continuously refine complex tasks. It supports production-grade workflows including live debugging, root cause analysis, financial modeling, and document generation across Word, Excel, and PowerPoint.
minimax
Speed-optimized variant of M2.7 with identical outputs, tuned for low-latency coding and agent tool-calling workloads.
minimax
MiniMax\\'s first multimodal release, accepting image and video input. Long-context inference runs cheaper and faster, and it handles multi-step agent workflows and computer-use scenarios.
Moonshot
Instruction-tuned for agent workflows with reliable tool calling and multilingual coding.
Moonshot
Open-source thinking model built for long-horizon reasoning and hundreds of sequential tool calls across autonomous research, coding, and agent workflows.
Moonshot
Top Chinese-English bilingual. Strong long context.
Moonshot
Latest multimodal model in the K2 series, built for long-horizon coding, code-driven UI/UX generation, and multi-agent orchestration.
Moonshot
Coding-focused variant in the Kimi K2 family. Accepts text and image inputs, with reasoning on by default. Suited for long-horizon coding and agentic workflows.
Moonshot
A multimodal reasoning model for complex coding, knowledge work, and long-running agent tasks. It works across large codebases, calls tools, and debugs, using images, logs, and test output to refine its output.
nvidia
Balanced powerful model.
OpenAI
Stronger instruction following and edit precision; better preserves faces and logos; faster generation.
OpenAI
Offering state-of-the-art visual quality, precise instruction following, and support for large-scale batch processing.
OpenAI
Stable code generation with predictable output.
OpenAI
Near GPT-4o performance at lower latency and cost, suited for high-frequency interactions, coding, and vision tasks.
OpenAI
Fastest GPT. Sub-second responses.
OpenAI
Balanced general-purpose model covering dialogue, coding, and multilingual tasks at reasonable speed and cost for assistants and routine agent workflows.
OpenAI
A lightweight model tuned for low-cost, low-latency text tasks like chatbots, summarization, and bulk classification.
OpenAI
GPT-4o Mini Transcribe is OpenAI's smaller, cost-efficient speech-to-text model built on GPT-4o Mini audio capabilities.
OpenAI
GPT-4o Transcribe is OpenAI's high-quality speech-to-text model built on GPT-4o audio capabilities.
OpenAI
GPT-4o Transcribe Diarize is an automatic speech recognition (ASR) model with built-in speaker diarization, meaning it associates audio segments with different speakers in a conversation.
OpenAI
Reliable all-rounder for everyday tasks.
OpenAI
ChatGPT product-line snapshot of GPT-5, tuned for natural multi-turn conversation and consistent tone rather than agentic reasoning workloads.
OpenAI
Fast and low-cost; for high-concurrency text and reasoning tasks.
OpenAI
Cheapest GPT. Built for high-volume simple jobs.
OpenAI
The deep-reasoning flagship in the GPT-5 family, always running at high thinking effort and slower to respond, suited for hard analysis or coding where accuracy outranks speed.
OpenAI
Successor to GPT-5 that scales reasoning depth to task difficulty, returning fast answers on simple prompts and slower deliberation on hard ones, fit for general assistant, coding, and tool-use work.
OpenAI
Rolling snapshot of the GPT-5.1 build powering ChatGPT, tuned for natural multi-turn conversation rather than heavy reasoning workloads.
OpenAI
Previous flagship. Strong reasoning at a lower price.
OpenAI
Snapshot pointer to the GPT-5.2 Instant model that powers ChatGPT, optimized for everyday chat, writing, and light coding.
OpenAI
Flagship reasoning model tuned for agentic coding and long-context analysis, with steadier instruction following and fewer hallucinations on high-stakes workflows.
OpenAI
Chat-tuned everyday model that talks more naturally and refuses less, with better factual accuracy on common questions.
OpenAI
High-quality coding with stable refactoring.
OpenAI
Lightweight variant in the gpt-5.4 family, tuned for low-latency, high-volume tasks like classification, extraction, and sub-agent execution.
OpenAI
Lightweight GPT-5.4 sibling tuned for coding, tool use, and agent workflows at roughly twice the speed of the 5.4 Pro tier.
OpenAI
Lightweight variant in the gpt-5.4 family, tuned for low-latency, high-volume tasks like classification, extraction, and sub-agent execution.
OpenAI
Flagship GPT-5.4 variant with chain-of-thought reasoning tuned for agentic coding and multi-step problem solving on high-stakes tasks.
OpenAI
GPT-5.5 natively supports video comprehension, speech emotion recognition, and autonomous tool calling, delivering omni-modal foundational capabilities for agent-based execution of complex tasks.
OpenAI
Reasoning-heavy GPT-5.5 tier tuned for long-horizon agentic coding and complex multi-step workflows, with image input.
OpenAI
GPT-5.5 natively supports video comprehension, speech emotion recognition, and autonomous tool calling, delivering omni-modal foundational capabilities for agent-based execution of complex tasks.
OpenAI
Entry point of the GPT-5.6 lineup. A small, low-latency workhorse for chat, tagging, and lightweight agent loops, keeping reasoning solid enough for routine work.
OpenAI
Top of the GPT-5.6 lineup, tuned for hard reasoning, multi-step coding, and agent runs that stretch over many turns. Well suited to CLI-driven development and problems that need to be worked through step by step.
OpenAI
The middle rung of the GPT-5.6 lineup, tuned for daily dev work, general reasoning, and agent orchestration. Delivers reliable performance at a manageable cost for routine business workloads.
OpenAI
Reasoning model that works through problems step by step before answering, useful for math, science, and code questions that need multi-step derivation.
OpenAI
Cost-efficient reasoning model with adjustable thinking effort, tuned for STEM and coding tasks where deliberation matters.
OpenAI
Reasoning-focused variant in the o3 family that produces longer chain-of-thought traces and accepts image inputs. Suited for harder math proofs, scientific reasoning, and multi-step coding tasks, with more consistent outputs across repeated runs.
OpenAI
Lightweight o-series reasoning model with built-in chain-of-thought, delivering steady math, coding, and tool-use quality at low latency and cost.
qwen
Instruction-tuned Qwen3 variant without thinking mode, strong on multilingual reasoning, math, code, and tool use for agent workflows.
qwen
Internal fine-tune of Qwen3-32B for text-only Chinese workloads, tuned to in-house instruction style for QA, summarization and rewriting.
qwen
Qwen3-generation thinking-mode reasoning model with native tool use, tuned for math, coding, and multi-step agentic workflows.
qwen
Deep-thinking preview that reasons step-by-step before answering, holding up on multi-step math, code, and Chinese-language tasks.
qwen
Instruction-tuned non-thinking chat model that answers directly without exposing reasoning, suited to agent workflows needing deterministic output like coding assistance and tool calling.
qwen
Lightweight multimodal model with balanced performance.
qwen
Plus-tier vision-language model in the Qwen3-VL line, handling text, image and video inputs with strengths in document parsing, video understanding, spatial grounding and agent tool use.
qwen
Open-weight mid-tier MoE text model with long chain-of-thought reasoning, strong on function-calling and multi-step planning benchmarks for self-hosted agentic workflows.
qwen
Mid-tier Qwen model tuned for multi-step reasoning and agent workflows, compact enough for private and on-prem deployment.
qwen
Lightweight mid-tier variant tuned for low-latency agent workflows with coding and tool-calling support.
qwen
Multimodal model from the Qwen3.5 line with toggleable thinking mode, tuned for image and video understanding, document parsing, and multimodal agents.
qwen
Thinking-mode variant of the dense open-weight model that outputs step-by-step reasoning; at 27B it surpasses the prior open-weight 397B-A17B flagship across coding, math and multi-step reasoning benchmarks.
qwen
Open-weight coding model tuned for agentic terminal tasks and repo-scale reasoning with low inference cost.
qwen
Qwen 3.6 lightweight tier with text, image, and video input. Low latency and low cost, well-suited for high-volume classification, extraction, summarization, and simple agent workflows.
qwen
Enhanced Qwen model with strong bilingual (Chinese/English) comprehension, excelling at long-document analysis and structured output.
qwen
Qwen3.7 Flash is the lightweight tier in the series, a vision-language reasoning model with strengths in object recognition, spatial understanding, and real-world visual perception. Suited for multimodal agents, visual coding, search, and computer interaction.
qwen
Flagship agent-centric reasoning model with a 1M context window, excelling at coding, productivity, and long-horizon autonomous tasks.
qwen
Qwen 3.7 Plus is a mid-tier multimodal model that reads screens, operates GUIs, and navigates mobile apps end-to-end. Suited for agent workflows and tool calling.
qwen
Flash tier in Qwen3.8: multimodal reasoning for coding and agent workflows, with visual understanding across documents, codebases, and long video.
qwen
Qwen3.8 series flagship, the general-availability successor to Max Preview. A multimodal reasoning model built for complex reasoning, visual understanding, coding, and agentic workflows.
stepfun
StepFun’s most capable open-source foundation model, built on a sparse MoE architecture that activates only 11B of its 196B parameters per token. It delivers efficient reasoning and fast performance, even with long contexts.
tencent
A 295B MoE reasoning model from Tencent with 21B active parameters, configurable reasoning effort, and a 262K context window. It is designed for agentic workflows, long-horizon tasks, coding, document processing, and financial analysis, with an emphasis on reliable tool use and reduced hallucinations.
tencent
Hy3 preview in high-reasoning mode, built for complex problem solving and multi-step agent workflows that need dependable execution.
xAI
Reasoning-oriented Grok variant built for multi-step problem solving and agentic tool use, with strict prompt adherence and low hallucination.
xAI
Multi-agent variant that dispatches sub-agents in parallel for deep research, multi-step reasoning, and tool orchestration.
xAI
Reasoning-oriented Grok variant built for multi-step problem solving and agentic tool use, with strict prompt adherence and low hallucination.
xAI
Grok 4.3 is a reasoning model from xAI that supports text and image inputs with text output. It is suited for agentic workflows, instruction following, long-document analysis, and deep research. The model supports a 1M token context window. Requests above 200K total tokens are billed at a higher rate.
xAI
Grok's flagship reasoning model, running an internal chain of thought before responding. Strong performance on coding and STEM tasks, with support for text, image, and file inputs. Suited for technical problem-solving and long-document analysis.
xAI
Grok's agent-oriented release, built for long-running multi-step work across a codebase. Accepts text, image, and file inputs. Suited for research and interactive or visual output.
xiaomi
Flagship agent model for complex workflows.
xiaomi
Mid-tier V2.5 model with native image, audio, and video understanding, built for agent workflows and coding at lower cost than Pro.
xiaomi
Flagship coding and agent model that sustains thousand-tool-call autonomous workflows for complex software engineering.
Z.ai
Text-only model in the GLM-4 family with improved coding benchmarks and tool calls during reasoning, suited for software development and agent workflows.
Z.ai
Multimodal model with strong image + text understanding.
Z.ai
Balanced model with strong Chinese understanding.
Z.ai
Ultra-fast for Chinese. Very cheap.
Z.ai
Competitive general-purpose Chinese model.
Z.ai
Designed for agent-based workflows such as OpenClaw.
Z.ai
Offers significantly improved coding and long-horizon task capabilities, autonomously planning, executing, and refining a single task for over eight hours to deliver complete, engineering-grade results.
Z.ai
GLM 5.2 is a large-scale reasoning model with a 1M-token context window, suited for long-horizon agents, repo-level coding, and multi-step automation.
Z.ai
Flagship-tier model for complex software engineering and long-horizon agent tasks, with stable execution across multi-turn tool calls and large-scale codebase refactoring.
Z.ai
Flash tier of GLM-5.3: native multimodal inputs, suited for efficient coding and long-horizon agents with stable long-context behavior in production.