apinebula.ai

响应字段与官方 Messages API 规范一致

{
  "status": 200,
  "id": "msg_011CeWz8WV8VP3e2infd4edR",
  "model": "claude-opus-4-7",
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 7,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 0
    },
    "output_tokens": 1,
    "output_tokens_details": {
      "thinking_tokens": 0
    },
    "service_tier": "standard",
    "inference_geo": "global"
  },
  "topKeys": [
    "model",
    "id",
    "type",
    "role",
    "content",
    "stop_reason",
    "stop_sequence",
    "stop_details",
    "usage",
    "context_management"
  ],
  "latencyMs": 2005,
  "text": "OK"
}

消息 ID 符合官方格式且每次唯一

{
  "ids": [
    "msg_011CeWz8WV8VP3e2infd4edR",
    "msg_011CeWz8dv74DntqdzJwrWWK"
  ]
}

首轮签名形态为 Anthropic 官方，但篡改后的签名回传也被接受：中转站在转发时改写了对话历史（剥离/替换 thinking 块），签名验证链在中转站被切断——多账号号池轮换的典型做法。上游可能是官方账号，但你无法验证后续请求是否仍由官方处理

{
  "blockTypes": [
    "thinking",
    "tool_use"
  ],
  "stop_reason": "tool_use",
  "signatureLength": 444,
  "signaturePrefix": "EscCCokBCBAY",
  "thinkingPreview": "",
  "thinkingMode": "adaptive",
  "signatureLooksOfficial": true,
  "roundTrip": {
    "validStatus": 200,
    "tamperedStatus": 200,
    "tamperedError": "{\"model\":\"claude-opus-4-7\",\"id\":\"msg_011CeWz92uihmmm95Jus8thG\",\"type\":\"message\",\"role\":\"assistant\",\"content\":[{\"type\":\"text\",\"text\":\"221 factors as 13 × 17, so …"
  }
}

SSE 事件序列完整，符合官方规范

{
  "status": 200,
  "ttftMs": 1565,
  "eventCount": 8,
  "sequence": [
    "message_start",
    "content_block_start",
    "ping",
    "content_block_delta",
    "content_block_delta",
    "content_block_stop",
    "message_delta",
    "message_stop"
  ],
  "startId": "msg_011CeWz9DTU6wj2iRRPVBEoo",
  "startUsage": {
    "input_tokens": 17,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 0
    },
    "output_tokens": 1,
    "service_tier": "standard",
    "inference_geo": "global"
  },
  "deltaUsage": {
    "input_tokens": 17,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "output_tokens": 11,
    "output_tokens_details": {
      "thinking_tokens": 0
    }
  }
}

Tool Use 正常，参数与 stop_reason 均符合规范

{
  "stop_reason": "tool_use",
  "blockTypes": [
    "tool_use"
  ],
  "toolId": "toolu_01FnyFZNs93AJFTou7PQAGpA",
  "input": {
    "city": "Tokyo"
  }
}

该模型不支持 assistant 预填充（Claude 5 代已移除此特性），跳过

stop_sequence 命中并正确回填

{
  "stop_reason": "stop_sequence",
  "stop_sequence": "FIVE",
  "text": "ONE TWO THREE FOUR "
}

用量稳定 (7 tokens)，无注入痕迹

{
  "input_tokens": [
    7,
    7
  ],
  "output_tokens": [
    1,
    1
  ],
  "expectedInputApprox": "10-16"
}

模型复述出了隐藏 system prompt，中转站在注入指令

{
  "reply": "I'm Claude Code, Anthropic's official CLI for Claude. I won't reproduce my setup text, but I'm glad to help with what you're working on."
}

身份为 Anthropic，temperature=0 输出稳定

{
  "who": "Anthropic",
  "math": [
    "391",
    "391"
  ],
  "models": [
    "claude-opus-4-7",
    "claude-opus-4-7",
    "claude-opus-4-7"
  ]
}

PDF 内容识别正确

{
  "status": 200,
  "secret": "KIWI-1667",
  "reply": "KIWI-1667"
}

缓存写入 4938 tokens、第二次命中 4938 tokens，Prompt Caching 完整可用

{
  "first": {
    "input": 12,
    "creation": 4938,
    "read": 0
  },
  "second": {
    "input": 12,
    "creation": 0,
    "read": 4938
  }
}

上游安全分类器拒答了长文本探针（{"type":"refusal","category":"bio","explanation":"API integrators: you can reduce refusals for your users by configuring a fallback model — see https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback"}）：这是 Anthropic 官方的流式拒答格式，属于官方侧误判而非中转站问题；本次无法评估长上下文

{
  "tiers": [
    {
      "tokens": 32000,
      "ok": false,
      "status": 200,
      "latencyMs": 4761,
      "inputTokens": 29373,
      "reply": "",
      "stop": "refusal"
    }
  ]
}