It might simply be that it was not trained enough in Elixir RL environments compared to Gemini and gpt. I use it for both ts and python and it's certainly better than Gemini. For Codex, it depends on the task.