AI TRENDS | Developers Find Workaround to Use Claude Code Without Anthropic Models

Developers are finding ways to run Claude Code with cheaper non-Anthropic models from rivals such as OpenAI. According to Sina Finance, the issue surfaced after Anthropic Claude Code lead Boris Cherny was drawn into an online dispute last month. Developer Alex Gateman wrote a proxy program that connected Claude Code to OpenAI's GPT-5.6 Sol model, and said his account was banned by Anthropic about 15 minutes after he submitted his first batch of coding requests on Claude. Claude Code is popular because it includes a full toolset that can read and run files, search code repositories, and execute multiple tasks in parallel. But the framework itself does not require Anthropic's own models, and many developers use third-party models to cut costs. Gateman's post drew more than 1 million views within hours. OpenAI's Codex lead Thibault Sottiaux said users should decide which model works best for them, while Cherny said Anthropic does not prohibit developers from using Claude Code with other vendors' models. Anthropic later restored Gateman's account and apologized privately. The company does not block proxy connections, but people familiar with the matter said its official view is that the tool works best with its own models and that self-built proxies can increase the risk of prompt-injection attacks. A spokesperson for Anthropic said the company has Claude models designed for different tasks, from high-difficulty work to cases where speed and cost matter most. The practice also reduces Anthropic's revenue because token fees go to the model provider being used behind Claude Code. OpenRouter competitor Not Diamond founder Thomas Hernando Kaufman said the approach is spreading among developers, and OpenRouter's public dashboard shows 366 different models have been connected to Claude Code so far. Adam Kosh, who is building an AI startup for manufacturing, said he uses OpenRouter and Concentrate to drive Claude Code with non-Anthropic models for industrial coding work. He estimated the setup preserves about 97% of output quality while cutting costs by about 60%.