Will you ever run Claude Code on GPT, or Codex on Claude, without losing performance?
retained text
Hot take: model-agnostic coding agents will always be second-class. Change my mind (with evidence) **I**'ve got this hypnotises that coding agents are becoming vertically integrated, the way Apple doesn’t run Android on iPhones. Claude Code is tuned for Claude, Codex for GPT, and Cursor is now part of SpaceXAI and training its own models with Grok. The harness (tool calling, context management, subagents, hooks, recovery loops) is co-optimised with each model’s quirks. So I think “bring any model” setups will always lose some performance. **The counter-case:** OpenCode and other open-source harnesses are model-agnostic, and Cursor has always let you pick your model. Maybe the harness turns into a commodity and model choice becomes a dropdown. **So which is it?** If you’ve run the same harness on different models, or the same model in different harnesses, what did you actually see? Benchmarks, failure modes, and enterprise procurement experience are especially welcome.
document fetched 2026-10-29 00:42 · canonical https://reddit.com/r/ClaudeCode/comments/1xi1dlj/will_you_ever_run_claude_code_on_gpt_or_codex_on