On 8 August I followed a public recipe for running OpenAI's model inside Anthropic's Claude Code. About fifteen minutes in, my Anthropic account was suspended. Before the first articles ran it was back, the leads of both products had argued about it in public, and within days the story was retold in six languages — mostly without the logs.

This is my version: with the logs, and with what I think about it now.

## Why I tried it {#why}

I test things constantly; it is most of what I do. At the time Claude Code was considered the best agent harness there was — number one. A ChatGPT subscription, meanwhile, can be used in other tools. So the question was obvious: how does GPT behave inside the best harness?

## The setup {#setup}

On 12 July Tibo Sottiaux, who leads Codex at OpenAI, posted a recipe for exactly that: install CLIProxyAPI, connect both Claude and Codex authentication, and point Claude Code at it. It ended with a promise — “If this gets blocked, I owe you a reset.”

I reproduced it almost exactly. Claude Code 2.1.226, unmodified. CLIProxyAPI 7.2.120, listening on `127.0.0.1:8317` and nowhere else. Codex signed in with OAuth.

## The suspension {#suspension}

The last request through the proxy finished at `16:46:49.366`. The first suspension email arrived at about 16:47 — seconds to tens of seconds later, and about fifteen minutes after my first request. The reason given was “suspicious signals”.

I was very upset. I did not expect the account to come back, and I certainly did not expect what came next. As far as I was concerned it was over. I filed an appeal and wrote a post more or less for the record — whatever happens, happens. I published the setup with it, and a warning not to use it until the cause was clear.

## What Anthropic and OpenAI said {#response}

The replies called Anthropic arrogant for pushing developers toward its own models. Tibo answered that users should decide for themselves which model is best for them.

Boris Cherny, who leads Claude Code at Anthropic, replied that Anthropic does not ban people for using other models in its harness, and that a different account classifier had almost certainly fired. Shortly after: “Unblocked.” The account came back — with a private apology, as The Information later reported.

Tibo kept his promise and reset usage limits for every paid ChatGPT Work and Codex user. Sam Altman's comment was one line: “lol, one of my favorite things about OpenAI is Tibo.”

## A correction: Luna, not Sol {#correction}

My first post said GPT-5.6 Sol, because that is what the recipe named. My local logs say the model that actually answered was GPT-5.6 **Luna**. Several write-ups still carry the first version.

## How far it went {#spread}

I knew it had gone big. That day all of AI Twitter seemed to be writing about it; for a day I was probably the most talked-about person in AI. What I did not know was how far it had travelled. I found out about a month later, when I sat down to do analytics: reported pieces in Japan and China, a reproduction with benchmarks, explainers, newsletters and reprints — sixteen write-ups in six languages.

Four newsrooms reported it, one developer reproduced it, and a dozen explainers and newsletters retold it. Most of the retellings say Sol, not Luna, and several headlines present the model choice as the cause, which nobody has shown.

| Outlet | Where | Date | Notes |
| --- | --- | --- | --- |
| [The Information](https://www.theinformation.com/newsletters/ai-agenda/developers-find-ways-use-claude-code-without-anthropic-models) | US · EN | 15 Sep | Reported; names the proxy and Sol |
| [PC Watch](https://pc.watch.impress.co.jp/docs/news/2131843.html) | JP · JA | 10 Aug | Reported; gets Luna vs Sol right |
| [36Kr · AI前线](https://eu.36kr.com/zh/p/3933059944775041) | CN · ZH | 10 Aug | Reported; says the cause was never confirmed |
| [CNMO](https://internet.cnmo.com/event/815483.html) | CN · ZH | 10 Aug | Reported; appeal, explanation, restoration |
| [Zenn](https://zenn.dev/yam2357/articles/claudex-proxy-codex) | JP · JA | 10 Aug | Reproduced claudex with benchmarks |
| [explainx](https://www.explainx.ai/blog/tibo-boris-claude-proxy-suspension-harness-freedom-august-2026) | EN | 9 Aug | Explainer; says Sol, warns the cause is unproven |
| [WindowsForum](https://windowsforum.com/news/claude-code-gpt-5-6-sol-ban-claim-remains-unverified.442134/) | EN | 9 Aug | Sceptical analysis: an allegation, not a policy |
| [We0 AI](https://we0.ai/articles/claude-code-gpt-5-6-sol) | EN · PT | 13 Aug | Long explainer; says Sol |
| [Pier99](https://pier99.substack.com/p/episode-6-anthropic-tried-to-hire) | EN | Aug | Newsletter; includes my correction |

Every other write-up, reprint and translation is on the [press page](/about/#press).

## The Information {#information}

On 15 September, five weeks after the ban, The Information published a reported piece on developers who run Claude Code with other companies' models. I worked closely with the reporter, Alix Coutures: calls, answering her questions, helping get the details right. My case is one of the piece's main examples.

It put on record what I had only seen in pieces: the account shut down about fifteen minutes after the first coding requests, posts that drew a million views within hours, and Anthropic's position — it does not ban people for using other models in Claude Code, but it does not support proxies either, because the tool is built to work best with its own models. It also reported that after my episode some developers turned away from the proxy, moving to routers such as OpenRouter instead, where Claude Code had been used with 366 distinct models by the day it ran.

It quoted me, too: “I tried it once, and I will never do that again.”

## Why I think it happened {#cause}

Honestly, I don't think Anthropic fully understands what its own bans are doing. There are a lot of false positives.

I understand where it comes from. Most of it is about China and distillation. Chinese labs distilled Claude heavily: they bought accounts on the $200 plan and resold them to people to make the money back. In February [Anthropic itself said](https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks) that three Chinese labs had run about 24,000 fraudulent accounts, many through services that resell access. Claude was afraid of that, and still is.

But the mechanism is opaque, and nobody outside knows what triggers it. I've looked: Europeans get banned, Americans sometimes, even corporate accounts. My guess is that Claude simply has too many users for anyone to review it all. If there is noise on Twitter, they look at your case. If there isn't, they may never look. Mine had plenty of noise.

So the honest answer is that I don't know why my account was suspended, and nothing public says. Nothing shows that choosing a non-Anthropic model caused it, and Anthropic has not published the notice, the classifier's result or an incident report. What the record supports is narrower: a suspension seconds after a localhost proxy session, a public explanation that pointed elsewhere, and a reversal before the week was out.

## What I do now {#now}

I don't run Claude Code on other companies' models any more. To be honest, I'm scared to: I don't want another ban. I stopped those experiments. There are a lot of experiments and only one of me.

For setup and billing references, see the [Claude Code guide](/claude-code/) and the [Codex guide](/codex/). The [Codex, Claude Code and Cursor comparison](/analysis/codex-vs-claude-code-vs-cursor/#trial) gives a method for testing coding workflows on the same project; it does not establish the cause of this suspension.

If you are a developer wondering what is allowed: in principle, anything. But Claude is overly cautious right now — because it is the leader, and because Chinese labs distil it. Don't play games with Claude. You will get banned, and the second ban will come easier than the first. Better not to joke with Claude. At least for now.