V2EX User Reports AI Coding Proxy Failure Suspected of Model Swapping, Recommends Official APIs

V2EX user selfly recently shared a frustrating experience with AI-assisted programming, highlighting the hidden risks associated with third-party model proxies. Due to the inconvenience of traditional git operations when frequently modifying files on remote machines, he planned to develop a web-based Git tool. After discussing requirements and forming a Markdown document on the web version of Doubao, he handed the task to opencode.

The first attempt used the combination of "opencode + proxy gpt-6.1-sol". After 3 hours of runtime and a cost of 7 RMB, the output interface was at a beginner level, and none of the core requirements such as fetch, pull, push, and functional areas were successfully implemented, described by the author as "pure waste". Given that gpt-6.1-sol should not perform so poorly, and lacking a ready GPT plan for direct comparison, he switched to testing deepseek-41-flash.

The result formed a stark contrast: similarly throwing the Markdown requirement file to the AI, the output was completed in about half an hour, with basic functions all usable and successfully implemented, exceeding the author's expectations. This huge gap in experience led selfly to infer that the problem likely lay with the proxy—so-called "deep water", implying that the proxy might have secretly replaced it with a cheaper or less capable model (such as the comment section speculation of 6luna acting as sol).

To avoid disputes, the author did not disclose the name of the proxy but noted its considerable popularity. This experience prompted his decision: to prioritize official APIs in the future. Although this case is a personal anecdote (n=1), it provides a specific warning for developers relying on AI agents for complex code generation: when output quality falls far below model baseline expectations, one must be wary of tampering or downgrading by the middle layer. Directly verifying model identity or returning to official channels is a viable path to avoid such engineering risks.