OpenAI Reveals a More Efficient Prompting Method for GPT-5.6

OpenAI has published a new set of prompting guidelines for GPT-5.6 Sol, which was released only a few days ago. The recommendations mark a major change in how users are expected to interact with AI models.

Previously, users were encouraged to write highly detailed prompts containing numerous rules, examples and step-by-step instructions. OpenAI is now offering nearly the opposite advice, encouraging users to avoid prompts that are unnecessarily long and complex.

GPT-5.6 Changes the Prompting Philosophy

According to OpenAI, GPT-5.6 performs best when prompts focus on the desired outcome, essential constraints, available data sources and completion criteria, rather than describing every step of the process in detail.

In launch documentation published on OpenAI Developers, the company said GPT-5.6 is significantly better than previous generations at independently selecting the appropriate reasoning approach, search strategy and tools.

As a result, lengthy instructions that may have been useful for GPT-5 can now become an unnecessary burden.

OpenAI recommends removing several common prompt components, including:

  • Rules that are repeated multiple times.

  • Writing-style instructions that do not affect the final result.

  • Examples that do not meaningfully influence the model’s behaviour.

  • Processing steps that GPT-5.6 can already perform reliably.

  • Tools or tool descriptions that are not directly relevant to the task.

Instead, prompts should retain only four core elements:

  • The goal.

  • Success criteria.

  • Mandatory constraints, including safety, policy or business requirements.

  • Stopping conditions.

This approach allows GPT-5.6 to independently choose the most effective way to complete a task instead of spending resources processing unnecessary instructions.

OpenAI also said GPT-5.6 follows prompts much more strictly than previous versions.

When a prompt contains overlapping or contradictory rules, the model may not simply ignore one of them. Instead, it may attempt to satisfy all the requirements simultaneously.

As a result, GPT-5.6 may consume more reasoning tokens, require additional processing time, increase costs and potentially produce lower-quality answers.

OpenAI therefore recommends removing duplicated or conflicting instructions before adding new requirements to a prompt.

Shorter Prompts Help AI Work More Efficiently

To test the new approach, OpenAI conducted experiments using its internal coding agents.

The results reportedly showed that simplifying system prompts alone increased model evaluation scores by approximately 10–15%, while reducing token consumption by 41–66% and cutting operating costs by 33–67%.

In other words, GPT-5.6 Sol can operate more efficiently and use fewer resources when it does not have to process excessive or redundant instructions.

OpenAI noted, however, that these figures came from internal testing and that real-world results may vary depending on the application.

Nevertheless, the findings point to a broader shift in prompting practices for GPT-5.6 models: prompts should be concise and focused while still clearly defining the goal, essential constraints and completion criteria.