Interactive ChatGPT answers should be judged by whether users finish tasks

The examples include a bill splitter, a savings calculator and a game. Those formats could save a user from turning a prose answer into a separate tool. A calculator that adjusts to new inputs, for instance, might make a planning task easier to complete. That is a plausible benefit, not proof that the tool gets the answer right or helps people understand it.

Aarush Selvan, an OpenAI product manager, told SiliconANGLE that the company aims to make complex topics easier to learn. The goal makes sense: for some tasks, an interactive diagram or calculator may communicate more clearly than a block of text. We should not dismiss that benefit just because it comes in a new interface.

But a polished layout can also make an answer feel more settled than its underlying evidence warrants. The relevant question is not whether a response looks engaging. It is whether users can reach a correct result, understand the information and finish what they came to do.

Test what people can do with the answers

OpenAI says GPT-6 uses a collection of prebuilt interface elements and software that can display parts of a response while the rest is still being generated. The company also says it trained GPT-6 to choose layouts and interactivity, and evaluated the resulting interfaces for clarity, usefulness and completeness. Those are OpenAI’s descriptions of its approach and evaluations; the announcement does not establish independent findings about user outcomes.

A meaningful comparison would hold the model and task constant, then test an interactive answer against a plain-text one. Evaluators should measure whether the answer is accurate, whether people complete the task, how well they understand the result and how long it takes to reach a useful outcome. For a calculator or chart, they should also check whether the output stays accurate when users change the inputs.

The evaluation should include cases where text is the clearer choice, as well as users with different accessibility needs. The announcement and reporting do not provide independent results on accessibility, comprehension or accuracy as people interact with the generated components. A feature that helps one task may add friction to another; the right answer depends on what happens in use.

Keep interface claims separate from model claims

The rollout adds another reason to distinguish the product’s presentation from the model’s capabilities. OpenAI said Intelligent UI applies to the Chat tab, not the models powering ChatGPT Work and Codex. It identified GPT-6 Sol for the initial paid-tier rollout and GPT-6 Luna for the scheduled Free and Go expansion. These details define where the change applies; they do not show what users can accomplish better.

OpenAI also said that GPT-6 Instant produces the first response to web-search questions, on average, 44% sooner than GPT-5.6 Instant. That company-reported comparison concerns when a response begins, not how quickly a task is completed. The announcement does not provide independent verification, and the figure does not by itself establish greater accuracy or capability.

A stronger interface could be valuable even if the underlying model were unchanged. And a more capable model would need to demonstrate its value beyond a more appealing presentation. Because the rollout combines both changes, broad impressions of the new ChatGPT experience cannot tell users which one produced a benefit.

We recommend that OpenAI commission and publish an independent, side-by-side user test of interactive and plain-text responses using the same model and tasks. The test should report accuracy, task completion, comprehension and time to a useful result, including how interactive outputs behave when users change their inputs. That evidence—not visual polish—should determine what claims about the new interface are justified.