People can definitely be productive with less powerful models. Supermaven or Cursor's tab autocomplete models from a year ago were already a huge boost over the pre-AI days. They just don't have the same capabilities as the leading models.
Curious if you've tried Gpt 5.4 High through Codex to compare for your use case?