chat 网页 · Platform API · OpenCode · OpenAI 兼容
DeepSeek 免费用,官方三个入口:chat.deepseek.com 浏览器/App 直接聊;platform.deepseek.com 注册拿 API Key,新号控制台通常显示约 500 万 token 试用额度;终端用 OpenCode(或其它 OpenAI 兼容客户端)连 `api.deepseek.com`。下文按场景写操作步骤。
试用额度不是无限,剩余量和有效期以 Platform 控制台为准(常见约 30 天)。网页聊天和 API 分开计费——写脚本、接 Agent 必须用 Platform Key。额度用尽后按官方价目充值,国内常见 支付宝 / 微信。
三种用法怎么选
|
场景 |
入口 |
要不要 Key |
|
试对话、写文案、长文总结 |
chat.deepseek.com |
不要 |
|
Python / Node 脚本、自动化 |
Platform API Key |
要 |
|
终端读仓库、改代码 |
OpenCode + DeepSeek 端点 |
要(Platform Key) |
网页聊天(不用 Key)
1. 打开 chat.deepseek.com 或官方 App2. 邮箱 / GitHub 登录3. 输入问题 → 回车4. 不用装 Node,不用 Key
适合试 prompt、日常问答、长文总结。高峰偶发 503,过几分钟再试。网页免费不能接程序——脚本和 Agent 一律走 Platform。
Platform API:拿 Key 并验通
# 1. 注册并创建 Key# platform.deepseek.com → API Keys → 创建 → 复制(sk- 开头)# 2. PowerShell 设环境变量(当前窗口有效)$env:DEEPSEEK_API_KEY = "粘贴你的Key"# 3. 在项目根目录验证python scripts/verify_deepseek_api.py# 成功会打印 OK# 4. OpenAI SDK 只改两处# base_url = https://api.deepseek.com# model= deepseek-v4-flash
接口 OpenAI 兼容,老项目通常只改 `base_url` 和 `model`。验证脚本走标准 `chat/completions`,返回 OK 即 Key 和模型名可用。模型优先 `deepseek-v4-flash`(快)或 `deepseek-v4-pro`(强);`deepseek-chat` / `deepseek-reasoner` 旧名官方已公告退役。思考模式用请求参数开,不靠 model 名。
终端 Agent:OpenCode 接入
# 前提:已有 Platform Key# macOS / Linuxcurl -fsSL https://opencode.ai/install | bashopencode/connect# 选 DeepSeek → 粘贴 Key# Windows:装 OpenCode Desktop,或 WSL / Git Bash 跑上面命令# 其它 OpenAI 兼容客户端$env:OPENAI_API_KEY = $env:DEEPSEEK_API_KEY$env:OPENAI_BASE_URL = "https://api.deepseek.com"
DeepSeek 文档列了 Claude Code、Copilot、OpenCode 等集成,端点都是 `api.deepseek.com`。OpenCode 适合读仓、改文件、跑 shell;长任务会快速消耗试用额度,跑前先看 Platform 剩余量。续费按价目页,Flash 输入约 1 元 / 百万 token(缓存未命中,以官网为准)。
和 Gemini 免费档比
|
|
DeepSeek |
Gemini AI Studio |
|
网页 |
chat.deepseek.com |
aistudio.google.com |
|
API 免费 |
新号约 500 万 token |
配额内 RPM 限额 |
|
终端 |
OpenCode + Key |
Gemini CLI 登录 Google |
|
上下文 |
v4 约 1M |
Flash 约 1M 级 |
Key 与常见报错
Key 当密码:用 `DEEPSEEK_API_KEY` 环境变量,别写进代码、别提交 Git。
402 余额不足:试用额度或充值余额用完 → Platform 看用量,按需充值。429 速率到顶:降并发或稍后再试。503(chat 网页):服务繁忙 → 换时段,或改走 API。400 模型名错误:先在 Playground / 文档确认可用 model 再写进脚本。
相关链接
DeepSeek 聊天
chat.deepseek.com
DeepSeek Platform
platform.deepseek.com
API 文档
api-docs.deepseek.com
Agent 集成说明
api-docs.deepseek.com/guides/agent_integrations
OpenCode 仓库
github.com/anomalyco/opencode
chat · Platform API · OpenCode · OpenAI Compatible
DeepSeek offers free usage through three official entry points: chat.deepseek.com for direct browser/app conversations, platform.deepseek.com for registering and obtaining API keys (new accounts typically show approximately 5 million token trial quota in the console), and via terminal using OpenCode (or other OpenAI-compatible clients) connecting to api.deepseek.com. Below are operational steps for each scenario.
The trial quota is not unlimited; remaining amount and validity period are determined by the Platform console (commonly around 30 days). Web chat and API are billed separately — scripts and agents must use a Platform key. After quota is exhausted, recharge according to the official price list, commonly via Alipay/WeChat in China.
Which usage method to choose
Scenario
Entry Point
Key Required?
Testing dialogues, writing copy, summarizing long texts
chat.deepseek.com
No
Python/Node scripts, automation
Platform API Key
Yes
Terminal reading repositories, editing code
OpenCode + DeepSeek endpoint
Yes (Platform Key)
Web Chat (No Key Required)
Suitable for testing prompts, daily Q&A, and long-text summarization. Occasional 503 errors may occur during peak times — retry after a few minutes. Web free mode cannot be connected to programs; all scripts and agents must use the Platform.
Platform API: Obtain Key and Verify
The interface is OpenAI-compatible; old projects typically only need to change base_url and model. Verification script uses standard chat/completions, returning OK indicates key and model name are valid. Prefer models like deepseek-v4-flash (fast) or deepseek-v4-pro (strong); legacy names like deepseek-chat/deepseek-reasoner are retired as per official announcements. Enable thinking mode via request parameters, not model name.
Terminal Agent: OpenCode Integration
DeepSeek documentation lists Claude Code, Copilot, and OpenCode integrations, all using api.deepseek.com as the endpoint. OpenCode is suitable for reading repositories, editing files, and running shells; long tasks consume trial quota quickly — check Platform remaining amount before running. Recharge based on price page, approximately 1 CNY per million tokens for Flash input (cached unhit, subject to official site).
Comparison with Gemini Free Tier
DeepSeek
Gemini AI Studio
Web
chat.deepseek.com
aistudio.google.com
API Free
New accounts ~5M tokens
Within quota, RPM limited
Terminal
OpenCode + Key
Gemini CLI with Google login
Context
v4 ~1M
Flash ~1M+
Key and Common Errors
Key as password: Use DEEPSEEK_API_KEY environment variable, do not hardcode or commit to Git.
402 Insufficient Balance: Trial quota or recharge balance exhausted → Check Platform usage, recharge as needed. 429 Rate Limit Exceeded: Reduce concurrency or try later. 503 (chat website): Service busy → Switch to API or try at off-peak hours. 400 Model Name Error: Confirm valid model name in Playground/docs before using in script.
Related Links
DeepSeek Chat
chat.deepseek.com
DeepSeek Platform
platform.deepseek.com
API Docs
api-docs.deepseek.com
Agent Integration Guide
api-docs.deepseek.com/guides/agent_integrations
OpenCode Repository
github.com/anomalyco/opencode

