Model Selection Guide
Browse by Type
Chat
150+ models, including GPT, Claude, Gemini, Qwen, and Doubao
AI Coding and Agents
Coding Agent benchmarks and selection guidelines
Image Generation
Text-to-image models such as Seedream and Gemini Nano Banana
Text to Speech
Text-to-speech (TTS) models
Speech to Text
Meeting transcription with ASR and speech recognition
Embeddings
Models that convert text into vector embeddings
Reranking
Models for semantic reranking
Web Search
Real-time search through the Web Search API
Document to Markdown
Convert PDFs, images, and Office documents to Markdown
Video Generation
AI video generation models
Define Your Selection Goals First
Do not start by choosing the model ranked first on an overall leaderboard. Narrow the field in this order:
- Hard constraints: Input and output modalities, tool calling, context length, API protocol, and available regions.
- Task type: General Q&A, complex reasoning, Chinese-language business tasks, code generation, repository-level Coding Agents, vision, audio, or video.
- Operational goal: Quality, cost, low latency, or high concurrency.
- Real-world validation: Run a representative set of tasks with your own prompts, tools, and data before choosing a default model and fallback.
An Agent's actual performance depends on the model, Agent harness, system prompt, tools, reasoning effort, and runtime budget. The same model can perform very differently in Codex, Claude Code, or OpenCode; a model-only leaderboard cannot tell you which Agent is best.
Third-Party Benchmarking Platforms
Leaderboards are useful for narrowing the candidate pool, but they are not a substitute for internal evaluation. Before using a result, check when the leaderboard was updated, which model version and reasoning effort were tested, which Agent harness was used, and how prices were calculated.
- General-purpose models: Start with Artificial Analysis Models to filter by quality, price, speed, latency, and context length.
- Coding Agents: Start with Artificial Analysis Coding Agents to compare complete model-and-Agent combinations from both Model and Agent perspectives.
These are good starting points. For Chinese-language tasks, subjective experience, repository engineering, front-end work, or terminal tasks, cross-check the specialized leaderboards below.
General-Purpose Model Selection
| Selection question | Recommended platform | How to use it |
|---|---|---|
| Which model's answers do users prefer? | Arena | A human-preference leaderboard based on anonymous head-to-head comparisons. Useful for subjective experience across Chat, Vision, Image, and Video, but not equivalent to factual accuracy. |
| Which model has more consistent objective capabilities? | LiveBench | Regularly refreshes questions and uses automated scoring across reasoning, mathematics, coding, data analysis, instruction following, and Agentic Coding. |
| How should I balance API quality, price, speed, and latency? | Artificial Analysis Models | Compare Intelligence, token prices, per-task cost, output speed, time to first token, and context length together. |
| How should I choose Chinese and domestic models? | OpenCompass | Typically updated every two weeks and covers open-source and commercial models, making it useful for reproducible comparisons of Chinese-language and academic capabilities. |
| How should I choose for Chinese industry-specific and multimodal tasks? | SuperCLUE | Review Chinese vertical leaderboards for software engineering, marketing, video, embodied AI, and other tasks. Check each sub-leaderboard's date and model coverage first. |
Coding Agents and Model Pairings
| Selection question | Recommended platform | How to use it |
|---|---|---|
| Which Coding Agent works best with the same model? | Artificial Analysis Coding Agents | Compare by Model or Agent across harnesses such as Claude Code, Cursor, and OpenCode, including success rate, cost, token usage, and execution time. |
| Which model is more reliable when real users work with tools? | Agent Arena | Based on real Agent sessions; focus on task completion, steerability, recovery after Bash errors, and tool hallucinations. |
| Which model is better at long-horizon repository engineering? | DeepSWE | Uses 113 original repository-level tasks and the same mini-swe-agent for every model, helping isolate model capability. |
| Which option is better for front-end and web generation? | WebDev Arena | Human blind evaluation of generated web pages, with filters for HTML, React, and different business categories; useful for evaluating visual quality and usability. |
| Can a complete terminal Agent finish multi-step tasks? | Terminal-Bench | Shows Agent and Model combinations together and is useful for comparing CLI, Shell, tool use, and end-to-end execution. |
Selection Guidelines
- Build a candidate pool: Use the Model List to identify models that meet your modality, API, context, region, and budget requirements.
- Choose a task-appropriate leaderboard: General Q&A, Chinese-language tasks, repository engineering, front-end work, and terminal tasks require different evaluations. Do not use one overall leaderboard for every scenario.
- Control comparison conditions: During A/B tests, keep the Agent harness, prompt, tools, reasoning effort, timeout, and maximum token budget constant.
- Record quality and operational metrics together: At minimum, record task success rate, human acceptance results, end-to-end latency, per-task cost, and failure categories.
- Retest with real business samples: External leaderboards only narrow the candidate pool. Base the final decision on your own codebase, business prompts, tool permissions, and quality standards.
- Reevaluate regularly: Model versions, prices, and leaderboards change. Retest production defaults and fallbacks on a fixed schedule.
How to Interpret Leaderboards Correctly
- Identify what is being evaluated: A leaderboard with a fixed Agent mainly compares models; one that varies both Agent and Model compares complete systems.
- Check confidence intervals before rank: When scores are close and intervals overlap, do not claim that one option is definitively better.
- Do not mix cost definitions: Token price, per-task cost, and subscription price are different, and none necessarily matches the actual Turing platform bill.
- Keep reasoning effort consistent:
medium,high,xhigh, andmaxcan significantly change quality, latency, and cost. - Do not copy static rankings: External leaderboards change quickly. This page explains where to look and which question each leaderboard answers.
- Return to your own tasks: Retest candidate combinations using your real codebase, business prompts, tool permissions, and quality standards, and prepare a fallback for production traffic.
This page does not recommend a default model. For candidate model specifications, pricing, APIs, regions, and availability status, see the Model List. To configure a Coding Agent, see the AI Coding Tools Guide.