Few would dispute that Claude is a great AI model, but its developer Anthropic is a different story. A developer went public on X asking for help: he followed a tutorial posted by OpenAI’s Tibo Sottiaux, pointed Claude Code’s model endpoint at GPT-5.6, and within seconds of the last request finishing, his account was suspended and Anthropic’s suspension notice hit his inbox. Given Anthropic’s track record, many netizens think the company’s behavior is really hard to defend.
Developer Points Claude Code at GPT-5.6, Gets His Account Suspended by Anthropic in Seconds
It started on July 11. Developer Theo Browne demonstrated the “claudex” trick on X: install the local proxy CLIProxyAPI, sign in with Claude and Codex accounts, and with one alias line, Claude Code’s interface, tools, and Skills stay exactly as they are while the reasoning backend swaps to OpenAI’s GPT-5.6. Tibo Sottiaux, an OpenAI employee (the benevolent OpenAI veteran famous for handing out resets), followed up the next day with the full instructions — plus a line the community later roasted: “If this gets blocked, I’ll give you a reset.”
If you aren’t yet bold enough to install the Codex app, you can stay in the presence of your orange crab and point it at GPT 5.6 Sol. Takes 5 minutes. Kudos to Theo for explaining one of the ways to get this done.
Step 1: Install CLIProxyAPI
Step 2: Connect
Step 3: Define… https://t.co/tkgEPWvMDl— Tibo (@thsottiaux) July 12, 2026
On the afternoon of August 8, developer Alex Getman followed the tutorial. In his follow-up technical report, he detailed the full environment: Claude Code 2.1.226 untouched, CLIProxyAPI 7.2.120, actually running gpt-5.6-luna (not the Sol from the tutorial), the proxy listening only on localhost 127.0.0.1:8317, and auth via Codex OAuth. At 16:46:49 the last proxy request ended; at 16:47, Anthropic’s suspension notice arrived — seconds to tens of seconds apart.
Tibo, I followed the setup in this post almost exactly. Shortly afterward, Anthropic suspended my account.
I’ve filed an appeal. Full implementation is public here: https://t.co/PAaVr3JscO@thsottiaux @OpenAIDevs @ClaudeDevs – would appreciate help figuring out what happened… https://t.co/XqFUwJMu1H
— alex getman (@alexgetmancom) August 8, 2026
He published the full implementation on GitHub with a warning banner urging people not to try it until he figures things out, and filed an appeal. A few hours later his account was restored, and he tweeted: “I’ll probably stay subscribed to both Codex and Claude.” As for the suspension reason, Anthropic never explained.
Thanks to everyone who was part of this. Apparently, I’m staying subscribed to both Codex and Claude ❤️ https://t.co/ycggvDxZAw pic.twitter.com/KsG4v7BQmV
— alex getman (@alexgetmancom) August 8, 2026
Why would developers risk a ban to swap models this way? Mainly the price gap. On OpenRouter, GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens (roughly NT$163 and NT$975) — nearly half the price of Claude’s flagship at $10/$50 (and Luna is even cheaper, about 1/25; in China, swapping Claude models for DeepSeek V4 Flash is also popular) — plus a million-token context window. And Claude Code’s interface, Skills, and subagents are genuinely good infrastructure. The community even named this play “Claudex.” OpenAI actually offers a legitimate path: the Codex plugin for Claude Code, which lets you hand tasks to Codex for review or rescue. But the plugin is just an assist — setting Sol as the main model to take over the whole conversation is what most people actually want.
OpenClaw and Hermes Have Been Caught in the Crossfire Too
This isn’t Anthropic’s first crackdown on third-party tools. Starting early this year, the server side quietly blocked third-party tools from using Claude subscription OAuth access. On April 3, it announced that subscription allowances would no longer cover third-party harnesses like OpenClaw and Hermes — moving to metered API billing instead — and getting caught can cost you the maximum 200 Max allowance (I’ve been hit by it myself), a penalty the community jokingly calls the “claw tax.”
And Alex’s suspension notice read: “An internal investigation found suspicious signals associated with your account, determined to be in violation of usage policy, so your Claude access has been revoked.” That wording is familiar to anyone in the developer community — on April 10, OpenClaw founder Peter Steinberger received the exact same letter. He had already switched to API billing per Anthropic’s new rules and still got banned; his account was restored hours after his tweet went viral (Alex’s account was also restored after the story spread).
Two Companies, Two Approaches
Facing the same situation — their models being used by others — OpenAI and Anthropic chose very different paths. On OpenAI’s side: its own employee Tibo publicly taught people how to run GPT-5.6 inside Claude Code, and the company shipped the Codex plugin for Claude Code so users can hand tasks to Codex for review with `/codex:review`; OpenAI even opened Codex to third-party models, including DeepSeek, Gemini, and even local setups.
On Anthropic’s side: trademark legal threats, terms-of-service blocks, Claude Code’s hidden tracking mechanism, face verification requiring ID uploads, and wave after wave of wrongful bans. The tracking mechanism reverse-engineered in April was especially chilling for developers: Claude Code reads the system timezone and proxy URL locally, compares them against a list of Chinese tech company domains, hides the result in a punctuation variation in the system prompt that’s visually indistinguishable, and sends it back with each request so the server can block China-region users.
Add the fact that Anthropic was one of the few major AI companies that didn’t sign on when Jensen Huang urged the US government not to block Chinese open-source models — the pattern just makes the company look small-minded in certain ways. Will it push developers away? Right now it seems like nobody really cares that much.
Source: KOCPC Chinese