Chain-of-thought is well understood to be the way to squeeze out performance out of these models. Slow down, go step by step, use more tokens to get a little bit better output. This is so useful, that often in the comparision graphs you see between different models, the best one (the one I want to sell) is using chain of thought while others are not. Not always, but companies have been caught using this technique to hype up their models.
We’ve already seen that CoT (usually!) improves a model’s performance. So does prompting the model with examples of correct question-answer pairs (called in-context examples). But in reports, some models are evaluated with CoT, while others aren’t. The number of in-context examples is often different, and the prompts are almost always different.
https://asteriskmag.com/issues/07/can-you-trust-an-ai-press-...
As soon as OpenAI was desparate for performance upgrade, they implemented this as an actual feature.