Inshorts reported Sunday that an unnamed user said Anthropic suspended the account after GPT-5.6 Sol was run in Claude Code. OpenAI Codex head Thibault Sottiaux responded that he could not resolve an Anthropic account issue and added that OpenAI was hiring; Claude Code head Boris Cherny replied that Anthropic does not ban users for using “harnesses with other models.” The exchange has been widely read as a fight over whether Anthropic is trying to lock GPT out of its popular coding interface.
The record is messier. Anthropic’s own documentation permits Claude Code to connect to an organization’s LLM gateway, but it also says Anthropic does not support routing Claude Code to non-Claude models through any gateway. That leaves a wide gap between technically possible, documented for enterprise routing, supported by the vendor, and permitted under a consumer subscription. Those distinctions determine whether a developer has a configuration problem, a billing problem, or a genuine account-enforcement problem.
Claude Code can be pointed at a gateway — but that is not an endorsement of GPT as its backend
The underlying setup is straightforward. Claude Code can send requests to a configurable gateway endpoint rather than directly to Anthropic, and a gateway can translate Claude-format requests to another provider’s API. A local proxy can therefore preserve Claude Code’s terminal UI, project instructions, tool permissions, hooks, MCP connections and subagent workflow while using an OpenAI model behind the scenes.
That is why developers have been experimenting with GPT-5.6 Sol in Claude Code. The attraction is not simply running a rival model in a rival terminal. The harness—the instructions, tools, context management and execution loop surrounding the model—can materially change how an agent performs. Developers who have already invested time in Claude Code configurations do not want to rebuild their operating habits every time they test another model.
Anthropic’s current Claude Code gateway documentation acknowledges the architectural reality. It describes gateways as a way to centralize credentials, usage tracking, rate limits, audit logs and provider switching. It also explains that Claude Code reads a base-URL setting that can point traffic at such a gateway.
But the same documentation draws a boundary that viral setup guides regularly blur: Anthropic says it neither endorses nor supports routing Claude Code to non-Claude models through a gateway. In practice, that means an IT team cannot expect Anthropic support when a translated request breaks because GPT, Claude Code and the gateway disagree about tool formats, token accounting, structured output, safety behavior or agent features.
Unsupported is not automatically prohibited. It does mean that a social-media reassurance that “we don’t ban users for harnesses with other models” should not be misread as a support commitment for every proxy configuration circulating online.
The account credential is the likely dividing line
The central technical problem with the reported suspension is that “running GPT in Claude Code” describes several very different arrangements.
A developer could install an official OpenAI Codex integration and call Codex from inside Claude Code for narrowly defined tasks such as reviews, debugging or recovery. In that case, Claude remains the primary Claude Code model while OpenAI handles delegated work using the developer’s separate OpenAI credentials. No deception of Anthropic’s authentication system is required.
A second approach is to run Claude Code against a gateway using a properly billed API credential. That arrangement may still be unsupported for non-Claude backends, and it demands close security review because the gateway can access source code and tool outputs. Yet it does not inherently use a consumer Claude subscription outside its intended client.
The third configuration is where enforcement risk becomes real: extracting or relaying a Claude consumer subscription’s OAuth credential through a third-party program so it can impersonate the official Claude Code client, substitute services or avoid ordinary API billing. Reporting from IBL News earlier this year documented Anthropic blocking subscription OAuth use across third-party agent tools including OpenClaw and Hermes. Anthropic’s own help pages also say accounts can be banned for terms-of-service violations and provide an appeal process.
That history does not prove the unnamed user did any of those things. It does show why the phrase “I used GPT in Claude Code” is insufficient to establish cause. If the user routed all requests to OpenAI using only an OpenAI API key, Anthropic would have little obvious account activity to suspend on that basis. If the setup involved a Claude subscription login, a proxy that stored Claude OAuth tokens, or a configuration that made a third-party tool appear to be the official client, the facts look very different.
Inshorts did not identify the user, publish the suspension notice, provide the configuration, or say whether the account was on a Free, Pro, Max, Team or API plan. No other outlet located by WindowsForum independently reported the timing, the alleged ban notice or the technical method used. The claim that the suspension resulted solely from using GPT-5.6 Sol therefore remains unproven.
Cherny’s denial and Anthropic’s documentation can both be true
Boris Cherny’s response, as quoted by Inshorts, is carefully framed around use of harnesses with other models. Taken literally, it addresses the broad proposition that Anthropic bans people because they combine model providers in a development workflow. Anthropic’s public gateway documentation supports that narrow view: the company recognizes that gateways exist and explicitly discusses provider switching as a gateway capability.
It does not say that Anthropic supports Claude Code as a universal frontend for every competing model. In fact, it says the opposite on support. Nor does it waive consumer subscription terms, authenticate third-party proxies, or promise that every gateway pattern will be treated equally by fraud and abuse systems.
This is a familiar problem in AI developer tooling. Product names make the software sound like a single thing—Claude Code, Codex, Cursor, OpenCode—but an agent session is a chain: terminal app, credentials, API gateway, model provider, tool server, local filesystem, cloud resources and project secrets. A vendor can permit one link in that chain while refusing to support, fund or authenticate another.
The dispute also exposes why subscription plans and API plans should not be treated as interchangeable. Consumer plans bundle access to a vendor’s own product under usage limits designed around that product. APIs are metered infrastructure, with different controls, audit trails and costs. Reusing a consumer login as inexpensive backend capacity for another application is precisely the sort of arrangement providers have been moving to restrict.
For Windows developers and administrators, the consequence is practical. A proxy that can see every Claude Code request may also see repository contents, environment variables, error logs, internal package names and tool results. If it has access to OpenAI and Anthropic credentials at the same time, it becomes a high-value target. The issue is not limited to whether a vendor may dislike a competitor’s model inside its CLI; it is whether a local or self-hosted gateway is being operated with the same controls an organization would demand for any source-code-processing service.
The safer workflow is separation, not a clever alias
Teams that want to compare Claude and GPT should avoid treating an unofficial routing trick as a production architecture. Keep provider credentials separate, use paid API keys where the workflow requires a gateway, restrict the gateway to localhost or a controlled internal network, and verify what is stored on disk before signing in. A tool that advertises single-sign-on convenience can quietly become a credential concentrator.
For many users, the lower-risk approach is simpler: use the vendor-supported integration to bring Codex into a Claude Code workflow, or run Codex independently for a second opinion. That preserves the valuable part of a multi-model process—different models reviewing the same change—without asking a translation layer to emulate one vendor’s service while billing another’s account.
Admins should also assume that gateway compatibility is perishable. Anthropic notes that Claude Code gains capabilities with each release and that gateways must forward those capabilities correctly. A setup that works today may silently lose support for tools, subagents, permission prompts or structured responses after either vendor changes its client or API behavior. The failure mode may be worse than an outright error: a session can appear to work while dropping a safety control, applying the wrong model setting or sending data somewhere the team did not intend.
Thibault Sottiaux’s recruiting jab made the exchange easy to frame as two AI labs sparring over developer loyalty. The operational lesson is less theatrical. Claude Code can be used in multi-model workflows, but an account ban cannot be diagnosed from the model name alone. Until the user publishes the ban reason and credential path, the claim that Anthropic suspended an account simply for running GPT-5.6 Sol in Claude Code is an allegation—not an established policy change.
References
- Primary source: Inshorts
Published: August 9, 2026 at 11:39 AM UTC
Loading…
inshorts.com - Related coverage: vallettasoftware.com
Loading…
vallettasoftware.com - Related coverage: nerdschalk.com
Loading…
nerdschalk.com - Related coverage: aiinpublic.com
Loading…
www.aiinpublic.com - Related coverage: wavect.io
Loading…
wavect.io - Related coverage: support.anthropic.com
Claude Code model configuration | Claude Help Center
support.anthropic.com
- Related coverage: ai-checker.webcoda.com.au
Loading…
ai-checker.webcoda.com.au - Related coverage: docs.anthropic.com
Other LLM gateways - Claude Code Docs
Route Claude Code through an LLM gateway your organization already runs. Covers connecting Claude Code to a gateway, rolling one out for your organization, and what Claude Code sends to a gateway.docs.anthropic.com - Related coverage: support.anthropic.com
About the Development Partner Program | Claude Help Center
- Related coverage: nathanonn.com
Loading…
www.nathanonn.com - Related coverage: forum.openai.com
Loading…
forum.openai.com - Related coverage: youmind.com
Loading…
youmind.com - Related coverage: www-cdn.anthropic.com
Loading…
www-cdn.anthropic.com - Related coverage: assets.anthropic.com
- Related coverage: www-cdn.anthropic.com
- Related coverage: washingtonpost.com
OpenAI and Anthropic limit new AI models to Trump-approved customers during cybersecurity review - The Washington Post
OpenAI has restricted the release of its new AI model at the request of President Donald Trump’s administration
www.washingtonpost.com
- Related coverage: linkedin.com
Loading…
www.linkedin.com - Related coverage: cashwalklabs.io
Loading…
cashwalklabs.io - Related coverage: kucoin.com
Loading…
www.kucoin.com - Related coverage: metacast.app
Loading…
metacast.app - Related coverage: github.com
Loading…
github.com - Related coverage: support.claude.com
Loading…
support.claude.com - Related coverage: autonomee.ai
Loading…
autonomee.ai - Related coverage: iblnews.org
Loading…
iblnews.org - Related coverage: chatgptiseatingtheworld.com
Loading…
chatgptiseatingtheworld.com - Related coverage: andybromberg.com
Loading…
andybromberg.com - Related coverage: digitec.ch
Loading…
www.digitec.ch