First understand the common terms of AI and AI providers
When you encounter Key, Base URL, model name, multiplier, and error codes, come here first to understand the terms before proceeding with integration and troubleshooting.
AI Basic Terms
| Noun | Plain explanation | What should beginners pay attention to? |
|---|---|---|
| model | The AI capabilities that are truly responsible for answering questions, writing code, and generating images or videos. | The model name must be copied exactly from the provider's backend; do not change the capitalization, hyphens, or version numbers yourself. |
| Large Model | A general term for models that can handle tasks such as natural language, code, images, etc. | More expensive does not necessarily mean it is more suitable for you; different models may be suitable for writing, programming, images, or videos. |
| Multimodal | The same model can handle different inputs such as text, images, audio, or video. | The tool supporting multimodal does not mean that your provider and the current model also support it. |
| Prompt / Instruction | The request you send to the AI, for example, "Help me summarize this content. " | The more specific the prompt, the easier it is to get usable results. |
| System Prompt | Rules with higher priority than ordinary questions, often used to constrain the role, style, and safety boundaries. | System prompts in code tools affect how it reads files and modifies code. |
| Context Window | The amount of content the model can remember and process at one time. | A larger context does not necessarily mean smarter, but it is more suitable for handling long documents, large projects, and continuous multi-turn conversations. |
| Automatic Compression / Automatic Summarization | When the conversation or file content is about to exceed the context limit, some tools will summarize the previous content into a shorter summary and then continue sending it to the model. | Automatic compression allows the conversation to continue, but the summary may miss details; when troubleshooting issues, modifying code, or checking configurations, do not assume the model still remembers all the original text. |
| Token | The basic unit for model billing and length calculation can be roughly understood as small fragments into which text is broken. | Both input and output consume tokens; long documents and long conversations will significantly increase costs. |
| Temperature | Parameter that controls the randomness of the response. | If you want stability and reproducibility, lower it; if you want divergent creativity, you can raise it appropriately. |
AI provider related
| Noun | Plain explanation | What should beginners pay attention to? |
|---|---|---|
| AI provider | Forward your tool request to one or more model service platforms. | It's not the model itself; the focus is on which models it supports, its stability, price, and risk record. |
| provider | A platform that provides AI interfaces, models, billing, and backend management, which can be official or an AI provider. | The API addresses, keys, model names, and billing methods of different providers may all be different. |
| Official channels | Services provided by the model vendors themselves, such as official platforms like OpenAI, Anthropic, Google, etc. | Stability and rules are clearer, but there may be thresholds for payment, region, price, and quotas. |
| Mirror / Proxy | Forward the request to the middle layer of the target service. | Pay special attention to privacy, stability, and whether request content is saved when seeing such terms. |
| New API / Secondary Development Panel | Many AI providers set up or further develop the backend based on open-source panels. | The menu name may be called Token, Channel, Grouping, Model Multiplier, and may not necessarily be the same as the official platform. |
| Channel | The line in the backend that connects to a certain upstream model service. | When the channel is abnormal, the Key may be correct, but the request still fails. |
| Model Grouping | Assign different models or different permissions to different groups. | When your key does not enter the corresponding group, the model will be unavailable or 'model not found'. |
| Model Mapping | The provider maps a display name to the actual upstream model. | The display name on the page may differ from the real model name; when configuring, follow the backend requirements. |
Common channel-group labels
Group names are usually provider-defined labels, not industry standards, and cannot prove the upstream source on their own. Before choosing one, verify the provider's documentation, supported clients, billing unit, quota limits, and actual verification results.
| Common label | What it usually means | What to confirm first |
|---|---|---|
| Channel group | A selectable group combining one upstream source with a set of pricing, permission, or rate-limit rules. | The same label can mean different things at different providers; rely on the current provider's explanation. |
| Official relay / direct API | Usually claims that requests are forwarded through the model vendor's official API. | The label itself is not proof; verify API behavior, model version, context, billing, and verification evidence. |
| Max / Pro / Plus / subscription pool | Usually pools quotas from official personal subscriptions or plans for shared use. | Check official rate limits, shared congestion, quota reset cycles, and whether only specific clients are allowed. |
| Claude Code / Codex only | A group accepting only requests from specific official CLI tools or clients, sometimes by checking protocol details, headers, or client characteristics. | Regular SDKs, third-party editors, or chat tools may not work. |
| Vertex / Bedrock | A route serving officially hosted models through Google Cloud Vertex AI or AWS Bedrock. | The model may be officially hosted, while authentication, regions, fields, and features can still differ from the vendor's direct API. |
| Kiro / Antigravity | Usually refers to routes converted from quotas offered by specific AI coding tools or promotions. | Price and availability depend on platform policy; quotas, client restrictions, and capabilities can change. |
| Reverse-engineered route | Wraps a web product or private protocol as an API instead of calling a public paid API directly. | Pay close attention to stability, privacy, context, streaming, tool calling, and terms-of-service risks. |
| Per-request billing | Charges by request count. It is a billing method, not an upstream-source label. | Confirm per-request input, output, and context limits; do not compare it directly with per-token pricing. |
Configuration Fields
| Setting | Plain explanation | How to Fill |
|---|---|---|
| API | The application calls the model's API. The tool does not open a web chat but sends the question to the provider via API. | Only required if the tool explicitly supports API configuration. |
| API Key / Key / Secret | A string used by the tool to prove "I am this account. " | Copy the full content from the provider's backend, do not take screenshots publicly, and do not copy extra spaces. |
| Token | What many AI providers call Key in their backend. | Seeing token, Token, API Key, or secret usually refers to the same thing: access credentials. |
| Base URL / API Address | The interface address provided by the provider for the tool to request the model. | It is the API link; sometimes it looks like a website homepage, sometimes it needs to end with /v1, follow the backend example. |
| Endpoint | Another name for the interface address. | If you see Endpoint in the provider documentation, usually look for the API Address or Base URL. |
| Provider | Provider configuration item in the tool. | Can be understood as "which provider, which Key, which API address I want to use. " |
| Model / Model ID / Model Name | Tells the provider which model's API to call. | Copy exactly according to the backend model list; manually add only if fetching the model list fails. |
| Compatible with OpenAI | The provider's API format imitates the OpenAI API, and many tools can simply fill in as OpenAI Provider. | Being compatible with OpenAI does not mean supporting all interfaces of Claude, Gemini, images, and videos. |
Billing and Limits
| Noun | Plain explanation | What should beginners pay attention to? |
|---|---|---|
| Balance | The amount or points you can still use in the provider account. | Insufficient balance can cause the request to fail, it is not necessarily a configuration error. |
| Pricing multiplier | A commonly used billing coefficient at AI providers, used to indicate how many times the price a certain model is charged. | The stronger the model, the larger the context, and the more complex the image and video tasks, the higher the multiplier may be. |
| Input price | The cost generated by the content you send to the model. | Long documents, long code, and long chat histories increase input costs. |
| Output Price | The cost incurred by the model's response content. | When having the model write very long content, the output cost will also increase. |
| Concurrency | How many requests can be initiated at the same time. | When concurrency is too low, opening multiple tools or sharing among multiple people will result in queuing or failure. |
| RPM / TPM | Requests per minute / Tokens per minute limit. | When reporting a rate limit, it usually means the requests are too fast or the content is too long. |
| Context length | How much content can one request carry at most? | When the limit is exceeded, the tool may truncate content, directly report an error, or automatically compress or summarize the historical conversation before continuing. |
| Cache | providers or models reuse already processed content to reduce costs or speed up. | Caching does not necessarily take effect automatically, nor is it necessarily completely free; it depends on the provider and model rules. |
| Cache price | A general term for cache-related costs; some back-end systems break it down into different prices such as creation, hits, reads, etc. | When seeing the cache price, first confirm whether it refers to write cache, read cache, or the backend merging several items for display. |
| Cache creation price / Cache write price | The cost incurred when writing a piece of reusable content to the cache for the first time. | When using long system prompts, long documents, or fixed context for the first time, this type of fee may be incurred. |
| Cache hit price / Cache read price | The cost incurred when subsequent requests reuse already cached content. | Usually lower than the normal input price, but may still charge; do not assume that hitting the cache is necessarily free. |
| Cache hit | A part of the content in this request was identified by the provider as the same as before and the cache was reused directly. | A cache will only be hit if it meets the provider's caching rules, such as having the same prefix, sufficiently long content, and being reused within the validity period. |
| Cache expiration time | The time range during which the cache can be reused. | Requesting again after expiration may incur charges as a regular input, or generate caching creation fees again. |
Common error words
| Error word | Common meaning | Priority check |
|---|---|---|
| 401 / Unauthorized | Authentication failed. | Whether the key is complete, expired, or has extra spaces copied. |
| 403 / Forbidden | The account or key does not have permission. | Model permissions, group permissions, regional restrictions, balance status. |
| 404 / model not found | The model does not exist or the current key cannot use this model. | Whether the model name is copied from the backend, and whether the Key is added to the corresponding model group. |
| 429 / rate limit | Requests are too frequent or the quota limit has been reached. | Concurrency, RPM, TPM, provider limits, please try again later. |
| timeout / timeout | The request did not return in time. | Base URL, network, provider status, whether the model is queued. |
| invalid api key | The key format or permissions are incorrect. | Copy the Key again, making sure not to confuse the remark name with the Key. |
| invalid base url | The interface address is incorrect. | Confirm that it is the API address, not the console address; is it needed? /v1 The back-end takes precedence. |
| insufficient quota | Insufficient credit or balance. | Account balance, model multiplier, whether the recharge has been credited. |
When you don't understand the error message, don't change many items at the same time. Check in order: Key, Base URL, model name, balance, model permissions, provider status.