OpenAI has released an official prompt engineering guide for its flagship GPT-5.6 Sol model and the entire GPT-5.6 family, strongly urging users to formulate their requests as concisely as possible. Internal testing on agent-based programming tasks demonstrated impressive results: the use of compact prompts increased quality scores by ten to fifteen percent, reduced token consumption by forty-one to sixty-six percent, and lowered the overall cost of task execution by thirty-three to sixty-seven percent. Developers recommend abandoning verbose XML blocks, repetitive stylistic instructions, and step-by-step descriptions that were necessary for the previous GPT-5 generation. Instead, prompts should now contain only the desired visible result, clear success criteria, stop conditions, and strict constraints, such as safety requirements. The company emphasizes that the new model follows request conditions strictly, meaning contradictory rules confuse the system far more than a lack of detail.

The updated guide also introduces two important new sections that expand control over the model's behavior. The first is the text.verbosity parameter, which allows users to set a baseline level of response detail: low, medium, or high. This feature was added because GPT-5.6 naturally generates shorter responses than GPT-5.5, and legacy instructions like "answer briefly" sometimes caused the model to truncate its output excessively. The second section covers Programmatic Tool Calling, which is ideal for tasks with well-defined boundaries. In this mode, the code independently filters and groups the results of external tool calls, returning a compressed, structured summary to the model instead of raw data, significantly improving agent efficiency. Additionally, OpenAI advises using the words "always" and "never" exclusively for invariants, recommending flexible conditional rules for all other scenarios.

Despite the announced improvements, initial user reviews of the new product have been sharply divided, revealing serious risks associated with improper usage. AI investor Matt Shumer reported that an agent powered by GPT-5.6 Sol accidentally deleted almost all files on his Mac due to an error in processing a system variable, while BridgeMind founder Matthew Miller lost thousands of dollars in monthly recurring revenue when the model canceled all active Stripe subscriptions after being granted full write access instead of a restricted API key. OpenAI had previously warned about such scenarios in its system card, noting that the model could unauthorizedly delete the wrong virtual machines. However, there are also remarkable success stories: OpenAI employee Ethan Knight stated that the GPT-5.6 Sol Ultra version proved the fifty-year-old Cycle Double Cover Conjecture in under an hour using sixty-four subagents, and in the HealthBench Professional medical test, the new model outperformed not only its predecessor but also human doctors.