光谱测速台 / SpectrumBench — Responses API 测速与用量

把速度、等待与额度，变成可复核的证据。
Turn speed, latency, and token usage into evidence you can audit.

简体中文 · English · 在线预览 · 部署指南 · 测量方法

简体中文

SpectrumBench 是一个本地优先的 Responses API 测速与额度观察台。它不追逐偶然峰值，而是锁定场景、推理、预热和缓存规则，以完整 usage 为准，分别呈现可见输出速度、端到端速度、首字延迟、缓存与 API 等效成本。

你会得到什么

• 三种固定单流场景：无缓存极限、日常连续对话、Codex 高缓存；不会把并发吞吐伪装成单请求速度。
• 实时看板、质量门禁、逐请求明细、历史可比性判断和一次性 JSON 报告。
• 完整 usage 与中断估算严格分开；预热会计入额度，但不进入正式速度分布。
• 鸢尾、青玉、晨曦三套浅色配色，以及背景精确为 #17191d 的全局暗色主题。
• Windows 一键脚本、本地 Python、Docker Compose 和 Linux systemd + SSH 隧道部署。

隐私与安全

API Key 和 Base URL 只在当前浏览器会话与服务进程内存中使用，不写入草稿、历史、报告或运行日志。服务默认仅监听 127.0.0.1；WebSocket 要求同源；远程上游必须使用 HTTPS；包含账号、查询参数或片段的接口地址会被拒绝。

这个工具会按你的操作向上游 API 发起真实请求，可能产生费用。自动化测试、Demo 和项目截图不调用真实模型。请先阅读隐私说明与测量方法。

本地启动

需要 Python 3.10 或更高版本。

python -m venv .venv
.\.venv\Scripts\python -m pip install -r requirements.txt
.\.venv\Scripts\python token_stress_test.py

Windows 也可以在安装依赖后双击 一键启动.bat；一键结束服务.bat 会先验证端口上的服务确实是 SpectrumBench，避免误结束其他程序。默认访问 http://127.0.0.1:18976。

Docker 与服务器

Copy-Item .env.docker.example .env
docker compose up --build -d

Compose 只把端口映射到宿主机回环地址。远程服务器请通过 SSH 隧道访问：

ssh -L 18976:127.0.0.1:18976 user@example.com

完整的 Docker、systemd、升级和备份步骤见部署指南。不要把含 API Key 的面板直接暴露到公网。

多端点命令行对比

为每个端点建立一个不会提交的 *.local.txt，依次写入 HTTPS 接口地址、API Key、模型名称三行，然后显式传入：

.\.venv\Scripts\python benchmark_token_speed.py `
  --configs endpoint-a.local.txt endpoint-b.local.txt `
  --mode daily-dialogue --samples 6 --warmups 1

程序不会自动寻找凭据文件。结果只记录配置文件名、模型和测量数据，不记录 Key 或 Base URL。

定价边界

仓库内置的是标准服务层、输入不超过 272K token 的静态 API 等效价格快照，用于统一换算，不代表套餐账单。超长上下文、Batch、Flex、Priority 和区域处理不套用这张表；真实价格以 OpenAI 官方定价为准。

English

SpectrumBench is a local-first Responses API speed and token-usage workbench. Instead of celebrating a lucky peak, it locks the scenario, reasoning level, warm-up policy, and cache rules, then uses complete usage data to separate visible output speed, end-to-end speed, time to first token, caching, and API-equivalent cost.

What you get

• Three locked, single-stream methods: uncached limit, daily conversation, and high-cache Codex. Concurrent throughput is never presented as single-request speed.
• A live dashboard, quality gates, per-request details, history comparability checks, and one-time JSON reports.
• Complete usage and interrupted estimates stay separate; warm-ups count toward usage but never enter the formal speed distribution.
• Iris, Jade, and Sunrise light palettes plus a global dark theme whose background is exactly #17191d.
• Windows launchers, local Python, Docker Compose, and Linux systemd + SSH tunnel deployment.

Privacy and security

The API key and Base URL exist only in the current browser session and service-process memory. They are never written to drafts, history, reports, or runtime logs. The service binds to 127.0.0.1 by default, WebSockets are same-origin, remote upstreams must use HTTPS, and endpoint URLs containing identity, query, or fragment data are rejected.

The tool sends real API requests only when you start a run, which may incur charges. Automated tests, the demo, and project screenshots never call a live model. Read the privacy note and measurement method first.

Run locally

Python 3.10 or newer is required.

python -m venv .venv
.\.venv\Scripts\python -m pip install -r requirements.txt
.\.venv\Scripts\python token_stress_test.py

On Windows, you can also run 一键启动.bat after installing dependencies. 一键结束服务.bat verifies that the listener is SpectrumBench before stopping it. Open http://127.0.0.1:18976.

Docker and servers

Copy-Item .env.docker.example .env
docker compose up --build -d

Compose maps the port to the host loopback interface only. On a remote server, use an SSH tunnel:

ssh -L 18976:127.0.0.1:18976 user@example.com

See the deployment guide for Docker, systemd, upgrade, and backup steps. Never expose the credential-bearing panel directly to the public internet.

Multi-endpoint CLI comparison

Create one ignored *.local.txt file per endpoint with three lines—HTTPS endpoint, API key, and model—then pass the files explicitly:

.\.venv\Scripts\python benchmark_token_speed.py `
  --configs endpoint-a.local.txt endpoint-b.local.txt `
  --mode daily-dialogue --samples 6 --warmups 1

The program never auto-discovers credential files. Results contain configuration filenames, model names, and measurements, but not keys or Base URLs.

Pricing boundary

The checked-in table is a static API-equivalent snapshot for the standard service tier with up to 272K input tokens. It is a comparison aid, not a subscription bill. Long context, Batch, Flex, Priority, and regional processing are outside this table; verify live rates on the official OpenAI pricing page.

Project map / 项目导航

Path	Purpose
stress_tool/	Measurement engine, FastAPI/WebSocket server, persistence allowlists, and dashboard
token_stress_test.py	Local/server launcher
benchmark_token_speed.py	Counterbalanced multi-endpoint CLI comparison
pricing_table.json	Auditable static API-equivalent pricing snapshot
tests/	Measurement, engine, privacy, URL, report, and server-security tests
docs/	Deployment, methodology, preview assets, and publication audit

MIT © ferretgeek