Has anyone benchmarked external OpenAI/Codex models in OpenClaude vs Claude Code behavior in Coding Agent CLI?

[FrorsttzNguyen]

I’m using 9router to connect my ChatGPT Codex subscription model into Claude Code.

It works, but subjectively it does not feel as good as using the model in its native environment. With the same repo and similar agent setup/planning files, the model seems less stable and sometimes less effective when routed through Claude Code.

I also occasionally see the CLI/session stop unexpectedly.

From what I understand, Claude Code’s original codebase and harness are primarily designed around Anthropic models, so there may be some limitations or assumptions when external/non-Anthropic models are plugged in through routers or OpenAI-compatible endpoints.

So I wanted to ask:

Has OpenClaude already improved this kind of external-model compatibility compared with Claude Code?

Specifically, I’m curious about things like:

• tool calling compatibility
• streaming/session stability
• context handling
• agent loop behavior with non-Anthropic models
• any Anthropic-specific assumptions that may affect external models

Has anyone benchmarked the same external model on Claude Code vs OpenClaude?

Thanks.