Anthropic banned a Claude Code account 15 minutes after it was pointed at OpenAI's GPT-5.6 Sol through a proxy

The Information reports that Anthropic shut down a Claude Code account within 15 minutes of the developer routing it to OpenAI's GPT-5.6 Sol through a code proxy. Reporter Alix Coutures writes that Boris Cherney, who heads Claude Code at Anthropic, found himself in the middle of the online spat that followed. Only three facts from the piece are used here because the article sits behind The Information's paywall.

What the developer did, and what happened next

Alex Getman built a proxy in code that let him type into Claude Code and have the requests answered by OpenAI's GPT-5.6 Sol. Fifteen minutes after his first coding request went in, Anthropic ended the account, according to the piece. Getman's posts about the ban drew a million views within hours, which is why the story became visible to anyone outside the small group actually running this setup.

The Information frames the underlying practice as a cost move. Developers are running Claude Code with cheaper non-Anthropic models to save money, since the tool ships with the parts a coding session needs (a file reader, a shell runner, a codebase search, a parallel task runner) and does not, as a matter of code, require Anthropic's own models to answer the questions.

Why the split matters

Claude Code as a piece of software has two independent halves that are easy to conflate. One is the harness: the parts that read files, run shell commands, search code, and coordinate parallel tasks. The other is the model that answers when the harness asks a question. The harness has value on its own, which is what Getman and others are trying to capture without paying Anthropic's model prices.

Anthropic's terms of service treat that separation as a boundary, and the ban is the first widely-visible confirmation that the enforcement is fast and automated enough to catch a proxy inside the first few minutes of use. Fifteen minutes is the timescale of a real-time telemetry check on outbound traffic, or a fingerprint on the payload shape that a model swap changes.

The read for a team on this stack

A team paying for Claude Code today has to make a call it did not have to make yesterday. Routing Claude Code through a proxy to a cheaper model is against Anthropic's terms, and Anthropic will now find that setup fast when it looks. A subscription that also feeds a team's day-to-day work can be lost before the finance win from the model swap covers a week's bills.

A legitimate version of the same idea is to run a coding harness that supports several model providers by design. Several open-source harnesses take that shape, and swapping models is what they are for. The piece does not name a preferred alternative, and this story is not the place for a comparison. What it does establish is that the pattern people were quietly doing until last week now has a public precedent for the outcome, and the outcome is the account.

Source

The Information: Developers Find Ways to Use Claude Code Without Anthropic Models by Alix Coutures, 15 September 2026. Getman's own post on the ban, linked from the article.