I had long heard that Claude is prone to banning accounts, and there are quite a few examples around me, so I had been using Codex before. But a while ago, I tried AnyRouter's free Opus 4.6, and together with the company's Claude API that I've been using during my internship recently, I found that Claude Code really is great. Subjectively speaking, Claude Code's tool calls feel more frequent than Codex's, the default permission management feels very comfortable to me, and it also feels better than Codex when handling all kinds of detail-oriented issues. However, since I don't use a Mac and don't use Codex Desktop either, I haven't really experienced the new features of the recent desktop version yet. (By the way, compared with Windows and Linux Desktop, I really dislike the logic of macOS. The Mac Mini I bought before is still sitting at home as a set-top box lol.)
So, with a GPT subscription and the recent GPT 5.5 being very powerful as well, what should you do if you want to use Claude Code? The answer is a CPA reverse proxy.
CPA Reverse Proxy
CPA, full name CLIProxyAPI, is used to aggregate and reverse proxy various AI APIs and subscriptions. After deploying CPA, you can use the same key and endpoint to access APIs for different models. For example, I have reverse-proxied:
- 2 GPT Plus subscriptions
- 1 Claude Pro subscription (be careful with this one; two accounts were banned recently, though I'm not sure whether the reverse proxy was the cause)
- A small amount of Grok API bought from the official site
- 2 free Gemini APIs
- Models within GLM's free quota
- Free models from OpenRouter
Deployment
Obviously, deploying CPA requires a server with a public IP, preferably overseas. (Of course, you can deploy it on a LAN, but then you won't be able to access it from multiple locations and devices, which doesn't work for me.) In addition, the latest version of CPA has removed the data statistics feature. If you need usage statistics and data persistence, you can use the cpa-usage-keeper plugin.
Also, if you are reverse-proxying Codex's GPT 5.5 for use in Claude Code, remember to set:
"autoCompactEnabled": true,
"autoCompactWindow": 258000
Because GPT 5.5 in Codex has a 400k context: 258k input, 128k output, and 14k reserved for compacting. Only the official API can unlock the 1M context. This also solved a previous confusion I had: when the context was relatively long, GPT 5.5 would throw an error, while GPT 5.4 would not. That is because GPT 5.4 in Codex supports a 1M context.
User Experience
- After using CPA as a reverse proxy, you can directly put CPA's key, Base URL, model, and other configurations into environment variables and enjoy GPT 5.5 in Claude Code. Turn on xhigh, rotate between two Plus subscriptions, and it's just incredibly smooth. So I always feel that what's powerful is not only Claude Opus, but even more so Claude Code.
- Besides using it in Claude Code, the reverse-proxied API can be used just like a normal API anywhere compatible with the OpenAI format. For example, this blog's AI spam detection and AI automatic translation both use the API reverse-proxied by CPA; a recent LLM-Mining project I'm working on also uses it.
Notes
Strictly speaking, reverse-proxying GPT and Claude subscriptions violates the terms of use. Recently, two Claude Pro accounts were suspended, though I'm not sure whether the reverse proxy was the reason. But then again, CPA's use of the built-in Codex and Claude Code access is theoretically not much different from a user accessing them directly. At least my two GPT Plus accounts have already used several hundred million tokens combined, and they have been very stable.