Maintenance Note
This is the series appendix, maintained long-term.
Inclusion rules:
- Only entries that ordinary people will encounter and that affect choices or safety boundaries
- Each entry under 250 words
- When products change, update facts first without rewriting series positions
- For model rankings, pricing, regional availability, and privacy policy — verify with official sources before writing
If you find outdated content, ask Codex to check the official page to confirm.
Models
GPT (Generative Pre-trained Transformer)
OpenAI’s large language model, powering ChatGPT and Codex. Latest versions are GPT-4o and the o1 series. ChatGPT web has a free tier; API is charged per token.
Claude
Anthropic’s large language model, known for long context and safety alignment. Claude 3.5 Sonnet is the current main version. Supports 200K context window, strong at complex reasoning and code tasks. Available at claude.ai and via API.
Gemini
Google’s large language model, native multimodal (text, images, audio, video). Gemini 1.5 Pro supports 1 million token context. Accessible via Google AI Studio and API.
DeepSeek
Chinese AI company DeepSeek AI’s models — DeepSeek Coder and DeepSeek Math series are known for high cost-performance. DeepSeek V3 and R1 are competitive with GPT-4 and Claude at significantly lower API pricing. Has web and API access.
Qwen (Tongyi Qianwen)
Alibaba’s large language model, strong at long Chinese context. Multiple versions iterated; latest is Qwen 3. Supports open-source model download, also available via API.
Kimi (Moonshot AI)
Chinese company Moonshot AI’s large language model, known for ultra-long context (up to 20M tokens supported). Kimi web and App available directly with a free tier. Very accessible for Chinese-language users.
Wenxin Yiyan (Baidu ERNIE Bot)
Baidu’s large language model, accessible via the Wenxin web interface or Baidu Intelligent Cloud API. Has a free tier; paid version offers higher limits.
MiniMax
Chinese AI company, main products are Hailuo AI (video generation) and general language models. Offers text, speech, video, and music multimodal models. API is pay-per-use.
Capabilities
Context Window
The maximum amount of text AI can process in a single input. Larger windows mean you can send more content at once. For example, Claude 3.5 Sonnet supports 200K tokens — roughly 150,000 Chinese characters or 500,000 English words. Check the tool’s context limit for long document tasks.
Multimodal
AI that can simultaneously process one or more of text, images, audio, or video. Examples: describing what is in an image, transcribing voice recordings, generating images. The supported modalities are a key dimension for choosing tools.
Web Search / Grounded Search
AI that can go online to query current information, answering from live data rather than only from training. Web search results typically include source links for verification. Not all AI tools support this — check before using.
Reasoning
AI’s ability to perform internal reasoning and thinking before giving an answer. OpenAI o1 and Claude 3.7 Sonnet are reasoning models with significantly stronger performance on math, coding, and logic problems — but with slower response times.
Memory / Context Recall
AI’s ability to remember previous content within the same conversation. Most tools’ “memory” is limited to the current session — cleared after closing the conversation. Cross-session memory requires additional tools (such as Codex’s MCP memory features).
How AI Works
Agent
A mode where AI autonomously breaks down tasks and executes multi-step operations. For example, “help me polish this draft and post it online” — an agent will break this down and execute steps in sequence without being directed step by step.
Tool Use / Function Calling
AI that can call external tools — read files, edit files, check weather, send emails, run code. Codex’s core capability is tool calling, enabling it to work in your actual directory.
MCP (Model Context Protocol)
Anthropic’s Model Context Protocol, a standard way for AI to connect to external data sources and tools. For example, Codex can connect to your Obsidian notes, file system, or other tools via MCP — more flexible than traditional API integration.
API (Application Programming Interface)
An interface for passing instructions between two software programs. When ordinary people use an AI API, the main task is providing an API Key for authorization, charged by token consumption.
Local Model
An AI model that runs on your own computer without internet access. Llama, Qwen, and DeepSeek all have open-source local versions. Good for privacy-sensitive use, avoiding subscription costs, or when customization is needed. Requires a sufficiently powerful GPU — or accepts slower CPU performance.
Costs
Token
The smallest unit AI uses to process text — roughly one Chinese character or a few English letters. AI charges by total input + output tokens. 1 Token ≈ 0.75 English words ≈ 1-2 Chinese characters.
Subscription
A fixed monthly fee with unlimited or capped usage. ChatGPT Plus ($20/month), Claude Pro ($20/month), GitHub Copilot ($4-10/month) are all subscription-based. Good for heavy daily users.
Pay-as-You-Go / Token-based
Charged by actual consumption, no monthly fixed fee. Good for users with variable usage patterns. OpenAI API, Claude API, and Google AI Studio all use this model.
Free Tier
The free usage allowance each tool provides. Usually daily or monthly limits, after which you pay or wait for reset. Most major tools have a free tier — good for testing and light use.
Safety
Hallucination
AI generating content that sounds plausible but is actually wrong. The cause is that the training objective makes it “fill in” what it thinks you want to hear rather than guarantee factual accuracy. Verify important facts with sources.
Privacy Policy
The tool provider’s policy on how they handle data you send. Policies vary widely: some explicitly do not use data for training, some retain usage rights, some offer enterprise data isolation. Read the privacy policy before use — or ask Codex to look up the official statement for you.
Training Data
The data used to train an AI model. If you do not want your data influencing AI training, choose a tool that explicitly states “user data is not used for training.”
API Key Management
An API Key is like an account password — if leaked, others have full account control. Principles: do not share it, do not commit it to public code repositories, do not store it in project files, use environment variables.
Interface Types
CLI (Command-Line Interface)
A tool running in the terminal window. Called “Terminal” on macOS and “CMD” or “PowerShell” on Windows. CLI can read local files, modify directories, and run commands — the best entry point for ordinary people using AI to work on local projects.
IDE (Integrated Development Environment)
A code editor that combines file browsing, terminal, and plugins. VS Code, Cursor, and Windsurf are all IDEs. Good for programmers or users who need to see file lists and chat side by side. Ordinary people using Codex CLI do not necessarily need an IDE.
MCP Server
The server side of Model Context Protocol — enables AI to connect to external data sources and tools. For example, Codex can connect to your Obsidian notes, file system, or other tools via MCP — more flexible and standardized than traditional API integration.
Public social links
