在 OpenAI Codex 里用 apimodels

三步开跑

1. 1

   拿一个 API key

   到 控制台 控制台新建一个 sk_… 字符串。

2. 2

   写 config.toml

   把下面这段贴到 ~/.codex/config.toml(没文件就新建)。

3. 3

   导环境变量 + 跑

   export 一下 key,直接 codex 起。

完整配置

# ~/.codex/config.toml — drop-in, paste as-is
model_provider           = "apimodels"
model                    = "gpt-5.4"          # or "gpt-5.5"
review_model             = "gpt-5.5"          # model used by /review
model_reasoning_effort   = "medium"           # low | medium | high | xhigh
disable_response_storage = true

[model_providers.apimodels]
name     = "apimodels"
base_url = "https://api.apimodels.app/v1"
wire_api = "responses"
env_key  = "APIMODELS_API_KEY"

# put your apimodels key in the env var the config points at,
# then launch codex (add this line to ~/.zshrc to make it stick):
export APIMODELS_API_KEY="sk_…your_key…"
codex

字段说明

推理努力等级

推理强度是请求体里的 reasoning.effort 参数(Codex 用 model_reasoning_effort 设置),不再是模型名后缀。所有等级单价相同;等级越高,reasoning tokens 越多(算进 output_tokens 一起计费)。直接调 API 时传 "reasoning": { "effort": "high" }。

30 秒自检

装 Codex 之前先用 curl 确认端点和 key 都能通:

curl -s https://api.apimodels.app/v1/responses \
  -H "Authorization: Bearer $APIMODELS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.4",
    "input": "Reply with exactly: ok",
    "reasoning": { "effort": "low" },
    "max_output_tokens": 16
  }'

期望响应(节选):

{
  "id": "resp_...",
  "object": "response",
  "model": "gpt-5.4",
  "status": "completed",
  "output": [{
    "type": "message",
    "role": "assistant",
    "content": [{ "type": "output_text", "text": "ok" }]
  }],
  "usage": {
    "input_tokens": 22,
    "output_tokens": 5,
    "total_tokens": 27
  }
}

看到 HTTP 200 + output[].content[0].text = "ok" 就稳了。

responses 还是 chat?

gpt-5-4 / gpt-5-5 现在两种都能调。Codex 用 wire_api = "responses"(它的原生模式,对推理和多轮工具调用处理得最干净),所以给 Codex 就选 responses。如果你的客户端只会发 OpenAI chat 格式,wire_api = "chat" 配 /v1/chat/completions 也能调到这两个模型。普通聊天模型(MiniMax-M2.5、Claude、Gemini 等)用 chat 即可。

排错