Codex uses chatgpt.com web endpoint instead of api.openai.com — causing 16,000+ connections/day and 65GB bandwidth in 6 days

Description

Codex desktop app (Windows) is routing all model inference through chatgpt.com (the ChatGPT web interface) instead of api.openai.com (the dedicated API endpoint). This is causing extreme bandwidth consumption and connection storms.

Environment

• Codex version: 26.602.71036 (prod build)
• OS: Windows 10.0.26200
• Config: model = "gpt-5.5", model_reasoning_effort = "xhigh"
• Network: Behind proxy (China → Japan VPN)

Evidence from packet logs (June 9, 2026 — single day)

| Endpoint | Connections | Purpose |
|---|---|---|
| chatgpt.com | 16,141 | Web interface — reconnects every ~3.6 seconds |
| chat.openai.com | 424 | API endpoint |
| api.openai.com | 1 | Actual REST API (barely used!) |
| ab.chatgpt.com | 458 | A/B testing (web-only feature) |
| persistent.oaistatic.com | 219 | Static web assets — reloads every 15 minutes |
| github.com (alive) | 2,034 | Normal heartbeat |

Bandwidth impact (6 days: June 7–13)

• Total download through proxy: 65.78 GB
• Estimated ChatGPT web traffic: ~55 GB (based on connection proportion)
• Normal API usage would consume ~2–5 GB/day, not 11 GB/day

Root cause analysis

1. Codex is connecting to chatgpt.com (web endpoint) instead of api.openai.com (API endpoint)
2. Web interface uses WebSocket/SSE with aggressive reconnect (1000+ connections/hour)
3. persistent.oaistatic.com reloads all static web assets every 15 minutes — this is unnecessary for a desktop app
4. ab.chatgpt.com A/B testing endpoint is being hit 458 times — desktop app should not participate in web A/B tests

Why this matters

I am in China and must use a paid proxy to access OpenAI services. This bug has cost me ~$40 in proxy bandwidth fees over 6 days for traffic that a properly-implemented API client would never generate. This is effectively a billing bug — I am paying for the ChatGPT web interface's overhead multiplied by Codex's connection behavior.

Requested fix

1. Make Codex use api.openai.com (or chat.openai.com) for model inference — NOT chatgpt.com
2. Eliminate or drastically reduce persistent.oaistatic.com asset reloads (desktop app should cache these)
3. Remove ab.chatgpt.com A/B testing from desktop app
4. Implement proper connection pooling — one connection every 3.6 seconds is not acceptable

Related

This may be related to issue #27911 (goals_1.sqlite write amplification: ~11 MB/s sustained writes) — suggesting systemic resource management issues in the current Codex build.