Cleaning up old GPT prompts can raise benchmark scores by 10-15% and slash token use by two-thirds

Anyone still running system prompts assembled over months of trial-and-error should take a close look at repetition. Per OpenAI’s own findings, stating the same instruction more than once does not reinforce it. It degrades it. Removing duplicated instructions raised benchmark scores by 10 to 15 percent and cut token usage by up to 66 percent, a meaningful efficiency gain at any call volume.

Delete instructions from your old prompts. OpenAI's rule is to state each instruction exactly once. They found that removing repeated instructions raised scores by 10 to 15% while cutting tokens by up to 66%. Nathaniel Whittemore

There is a second, related trap for teams migrating from GPT-5.5. GPT-5.6 defaults to shorter answers out of the box, which means brevity rules written for the older model can now “cut too much,” as Nathaniel Whittemore put it. An instruction that once reined in verbose output can silently strip useful content when applied to a model that is already more concise by default.

Both issues share the same root cause: prompts accumulate instructions that made sense at the time they were written, then go unaudited as the underlying model changes. The practical takeaway is straightforward. Each instruction once, and review any brevity or length constraints against the specific model version currently in use.