Compare models with a purpose

Start with the question you need answered

Model comparison is most useful when it helps a real decision. Before opening two response panels, decide what a useful answer looks like. For a writing task that might mean clear structure and faithful use of the source. For code it might mean a correct fix, relevant tests and an explanation you can check.

Give both models the same prompt. Include necessary context, the intended reader and any constraints. If one model receives a clearer brief than the other, the comparison tells you less about which model fits the task.

Look at substance before style

A confident voice can hide an unsupported claim. Compare whether each response answers the actual question, identifies missing information and separates facts from assumptions. Check references yourself. Two models can repeat the same mistake, so agreement is not independent verification.

For code, run it in a suitable environment before accepting it. For calculations, inspect the inputs and units. For a document summary, compare the answer with the original text. The useful part of a comparison is often the question it prompts you to investigate next.

Keep the conditions visible

Max Harness labels responses with their provider and model. Generation settings can affect the output, and not every model accepts the same controls. A provider may change a model behind an alias. If a result matters, record the model ID and the date alongside your working notes.

Remember that comparison creates two provider requests. Both may use credits, and the same prompt is sent to both selected services. Consider the content and provider data policies before comparing sensitive work.

Follow up where it helps

After comparing, continue in the conversation whose answer gives you the most useful starting point. Ask about a specific gap or challenge a claim. A second response to a better question often helps more than collecting a large number of first drafts.

Use projects to keep related comparisons with the rest of the work. If you find yourself repeating an instruction, make it part of the project or a reusable assistant. Keep the instruction concrete so it improves the next request.

Learn about model settings or open the workspace to try a comparison with your own connected providers.