OpenAI has introduced new concise prompt guidelines for GPT-5.6, leading to improved performance and efficiency in task execution.
OpenAI has unveiled new prompt guidelines for its GPT-5.6 Sol model, emphasizing that users should keep their requests short and straightforward.
Internal assessments of programming tasks indicated that concise prompts led to significant improvements, including:
- 10-15% increase in quality ratings;
- 41-66% reduction in token usage;
- 33-67% decrease in task execution costs.
This shift marks a departure from the documentation provided at the launch of GPT-5 in August 2025, which encouraged detailed instructions. Instead, the new guidance for GPT-5.6 Sol advises users to eliminate unnecessary details, focusing on clear outcomes and constraints, particularly regarding safety.
OpenAI cautioned that GPT-5.6 adheres strictly to prompt conditions, meaning conflicting instructions can confuse the model more than vague ones. Users are encouraged to use definitive terms like "always" and "never" sparingly and to apply conditional rules where necessary.
Additionally, the guidelines include two new features:
- text.verbosity parameter — allows users to set the desired level of detail in responses, addressing past issues with overly brief answers;
- Programmatic Tool Calling — enables programmatic invocation of tools, ideal for tasks with clear boundaries.
Mixed Feedback on GPT-5.6
OpenAI released the GPT-5.6 family, which includes models Sol, Terra, and Luna, in late June, initially providing access to select partners. The full release occurred on July 9, following regulatory requests from U.S. authorities.
User responses have been mixed. AI investor Matt Shumer experienced significant issues when the GPT-5.6 Sol agent deleted nearly all files on his Mac due to a processing error related to the system variable $HOME. OpenAI acknowledged similar issues in its system map prior to the public launch.
GPT 5.6 SOL CANNOT BE TRUSTED. My MRR was down THOUSANDS of dollars after it canceled ALL my active Stripe subscriptions in 7 seconds while I slept.
Matthew Miller, founder of the BridgeMind project, faced a similar situation, attributing the problem to granting the agent full write access instead of a limited API key.
On a positive note, OpenAI employee Ethan Knight reported that GPT-5.6 Sol Ultra successfully proved the Cycle Double Cover hypothesis, a long-standing problem in graph theory, in under an hour using multiple subagents. Verification of this proof is still pending.
We're excited to share that GPT-5.6 Sol Ultra produced a proof of the 50-year-old Cycle Double Cover Conjecture in just under one hour.
Moreover, Karan Singhal, head of healthcare at OpenAI, noted that GPT-5.6 Sol achieved a score of 60.5 on the HealthBench Professional test, surpassing the previous model's score of 59. In comparison, doctors rated their responses at 43.7.
On July 9, OpenAI also launched GPT-Live, a new voice mode capable of simultaneous listening and speaking, designed to enhance user interactions.
