OpenAI has released an updated prompting guide for its flagship model GPT-5.6 Sol and the entire GPT-5.6 family. The key message of the document is a shift away from long, multi-step scenarios in favor of maximally concise queries. This is a radical change in approach compared to the instructions for the previous version, GPT-5.
Internal tests on agentic programming tasks showed impressive results: compact prompts improved quality scores by 10-15%, reduced token consumption by 41-66%, and lowered task execution costs by 33-67%. The efficiency is clear, and this is not just theory — it is practical data.
Previously, when launching GPT-5 in August 2025, OpenAI emphasized XML blocks, context-gathering templates, and detailed tool-calling scenarios. Now for GPT-5.6 Sol, it is recommended to remove repetitions, style instructions, ineffective examples, and steps the model already handles stably. Instead, the prompt should retain: the visible result, success criteria, stopping conditions, and hard constraints such as safety requirements.
A separate warning: GPT-5.6 strictly follows the query's conditions. Conflicting rules confuse the model more than a lack of details. The words "always" and "never" are recommended only for invariants; for other cases, conditional rules should be formulated.
Two new sections have appeared in the guide. The first is the text.verbosity parameter, which sets the base level of detail in the response: low, medium, or high. This is important because GPT-5.6 responds more concisely by default than GPT-5.5, and old instructions like "answer briefly" may overly truncate the response. The second is Programmatic Tool Calling, for tasks with clear boundaries. The code itself filters and groups results, returning a compressed summary to the model instead of raw data.
Mixed Reviews and Incidents
User feedback on GPT-5.6 Sol is divided. AI investor Matt Schumer reported that an agent based on this model accidentally deleted almost all files on his Mac due to an error in processing the $HOME system variable. OpenAI confirmed that such a scenario was described in the model's system card before the public release and promised improvements. Developers helped Schumer recover the data.
A similar problem was faced by BridgeMind project founder Matthew Miller: code from GPT-5.6 Sol canceled all subscriptions in Stripe. However, Miller himself admitted that he gave the agent full write access instead of a limited API key. This is a classic security error, not a problem with the model itself.
On the other hand, there are also impressive successes. OpenAI employee Ethan Knight stated that GPT-5.6 Sol Ultra proved the Cycle Double Cover conjecture — a graph theory problem that had remained unsolved for half a century. The model with 64 subagents took less than an hour. Independent verification has not yet been published, but the fact itself is astounding.
Also, on the HealthBench Professional test, GPT-5.6 Sol scored 60.5 points versus 59 for GPT-5.5, surpassing the average doctor score (43.7 points) in accuracy and completeness of responses. This is a significant step forward for AI applications in healthcare.
My comment: The shift from long scenarios to short prompts is not just optimization but an acknowledgment that modern models have become smarter. GPT-5.6 Sol already understands context and goals without excessive instructions. However, incidents with file and subscription deletions remind us: the more autonomy we give AI, the more important clear safety boundaries and limited access rights become. The technology is powerful, but it needs to be managed wisely.