Key facts
- OpenAI's new GPT-5.6 Sol prompting guide emphasizes outcome-first prompting.
- Leaner system prompts improved coding agent evaluation scores by 10-15% in internal tests.
- The new guidelines reduce the need for extensive system prompts, style rules, and examples.
- GPT-5.6's reasoning tokens are consumed when reconciling conflicting rules, leading to slower and more expensive outputs.
- New features include the text.verbosity parameter and Programmatic Tool Calling.
OpenAI has introduced a new prompting guide for its latest flagship model, GPT-5.6 Sol, which marks a significant shift from previous advice. The core message is to reduce the length and complexity of system prompts, adopting an 'outcome-first' approach. This means defining the desired outcome, setting stopping conditions, and allowing the model to execute without excessive instruction.
Internal testing by OpenAI demonstrated that leaner system prompts led to an improvement in evaluation scores of approximately 10-15% for coding agents. Furthermore, this approach resulted in a substantial reduction in token usage, between 41-66%, and a corresponding decrease in costs, ranging from 33-67%.
The previous GPT-5 prompting guide, released in August 2025, focused on adding 'scaffolding' such as XML persistence blocks and detailed context-gathering templates to guide the model's behavior. In contrast, GPT-5.6 largely negates the need for such extensive scaffolding, as the model is now better equipped to handle tasks without explicit, step-by-step instructions. The new guide advises users to retain only the essential elements: the user-visible outcome, success criteria, stopping conditions, and hard constraints.
OpenAI warns that GPT-5.6 is sensitive to prompt contracts, and conflicting rules can lead to instability, increased processing time, and higher costs as the model attempts to reconcile them. The company advises against using absolute commands like 'always' or 'never' to steer behavior.
Two key additions in the GPT-5.6 guide are the text.verbosity parameter, which allows for default conciseness settings, and a new section on Programmatic Tool Calling. This feature is designed for bounded workflows where code can manage filtering and aggregation of intermediate outputs, offloading complex processing from the model.