Choosing an AI Model

The model selector determines which AI handles your next request. Choose it before sending a build or review, and check it when reopening an older project. Projects can retain a saved model while Model Settings controls the generations offered for general use.

Models differ in cost, response style and supported inputs. A clear specification gives you a better basis for assessment than repeatedly sending a broad request to different models. Keep the same test case when comparing results so you can judge whether the program improved.

Match the model to the problem

Use the available coding models for new programs and well-defined changes. Before changing model for a difficult problem, provide complete source, exact diagnostics or a reproducible example. A model still needs the missing information about trading rules and dependencies.

GPT Codex 5.3 is offered for user-controlled continuation when automatic compiler repair stops. Continue with Codex prepares that workflow with source and diagnostics. Review them before proceeding, and test the returned behaviour even if compilation succeeds.

Images and model generations

Reference images require a supported model. The workspace disables image upload for GPT Codex 5.3 and indicates that Sol, Luna or Claude can be used instead. Check the selection when an image is already attached rather than assuming every model can read it.

Model Settings switches between available latest and previous generations. Use the options currently shown because names and availability can change. A saved project can retain its earlier model until another generation is applied; confirm the dropdown after changing that preference.

Compare cost with useful results

Usage depends on source size, accumulated conversation, generated output and the model. Focused changes and tested snapshots can reduce unnecessary regeneration. Keep closely connected source together when removing context would make the change harder to perform correctly.

Judge the result through compilation and behaviour tests rather than the confidence of its explanation. If responses interpret a rule differently, clarify that rule first. See Credits and Account for balance management and Existing EA for repair controls.