Test MiniMax M3 inside Claude Code with three real coding tasks

Connect the exact route from the recording

Thanks to MiniMax for sponsoring this video.

Luke changes the Claude Code settings.json configuration to use the MiniMax API. Use the retained platform and documentation links for the actual setup rather than copying invented endpoint values.

1. Inspect the model and account requirements at the source destination.
2. Keep a backup of the working client configuration.
3. Follow the recorded API setup without exposing credentials.
4. Verify the selected model route before assigning repository work.

Define the same three jobs and acceptance checks

The video tests a landing-page redesign, a production-bug repair and an invented feature on a real SaaS. A model preference should remain attached to those specific tasks.

1. Save the input brief and starting code for each comparison.
2. For the page, inspect clarity and actual interface behaviour rather than assuming CTR improved.
3. For the bug, keep a reproduction that fails before the repair and passes afterward.
4. For the feature, define the user task and inspect whether the implemented behaviour meets it.

Compare outputs, usage and effort

Luke reports M3 winning all three of his runs against Opus 4.8. The source does not establish a universal ranking or a measured conversion uplift.

1. Review the actual diffs and completed tests for each result.
2. Record follow-up instructions and manual fixes needed.
3. Compare usage, cost and time on the same basis.
4. Keep the July model and plan context attached to any summary of the verdict.

The takeaway

The useful comparison is reproducible work on the same tasks. A title’s win claim should not be expanded beyond the recorded tests.