Is Codex getting worse?

We look for the same downward movement across multiple reasoning efforts. One lower score can be ordinary benchmark variation, so the page only calls out a possible broad drop when at least two comparable efforts fall together.

Sometimes. A higher reasoning setting is not automatically better for every task. The effort comparison shows the latest measured score for each setting, and leaves a setting blank when no public result exists.

No. Juice is an observed runtime-budget signal. Capability scores come from coding-task results. A Juice change may tell us that a configuration changed, but it does not prove that coding quality improved or declined.

Yes. Select GPT-6 Astra to view its public benchmark observations when available. GPT-5.6 Sol retains its own history. Trends compare only the same model, reasoning effort, and benchmark suite; a model launch alone does not establish an improvement or decline.