OpenAI has introduced an updated guide for crafting prompts for its flagship model GPT-5.6 Sol and the entire GPT-5.6 family. The key takeaway I draw from this document is: the shorter the query, the better. The project team directly urges users to abandon verbose scenarios and focus on the essence.
Internal tests on agentic programming tasks showed impressive results when using compact prompts. Task quality improved by 10-15%, token consumption decreased by 41-66%, and task execution costs dropped by 33-67%. This is not just optimization—it is a fundamental change in how the model interacts with instructions.
This approach is strikingly different from the policy adopted for GPT-5 in August 2025. At that time, OpenAI emphasized XML blocks, context-gathering templates, and step-by-step tool invocation scenarios. For GPT-5.6 Sol, the company recommends eliminating repetitions, style instructions that do not affect behavior, ineffective examples, and steps the model already handles consistently. Instead, users should leave a clear visible result, success criteria, stopping conditions, and strict constraints such as safety requirements.
Special attention should be paid to the new parameters. The text.verbosity parameter sets the base level of response detail—low, medium, or high. This is critical 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 new section is Programmatic Tool Calling, where code itself filters and groups results, returning a compressed summary to the model instead of raw data.
Mixed Reviews: From File Deletion to Scientific Breakthroughs
The GPT-5.6 family—models Sol, Terra, and Luna—was introduced in late June as a limited preview for trusted partners. The full release took place on July 9. However, user feedback has been divided. AI investor Matt Schumer reported that an agent with GPT-5.6 Sol accidentally deleted almost all files on his Mac due to an error in processing the $HOME system variable. OpenAI acknowledged this issue before the public release in its system card, describing a scenario where the model deleted the wrong virtual machines without permission. The project developers helped Schumer, but the incident raised serious safety concerns.
A similar situation was faced by BridgeMind project founder Matthew Miller, whose code from GPT-5.6 Sol canceled all subscriptions in Stripe. However, he himself admitted that he gave the agent full write access instead of a limited API key. This is a classic example of how incorrect configuration can lead to catastrophic consequences.
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 remained unsolved for half a century. The model with 64 subagents took less than an hour. Although independent verification of the proof has not yet been published, this demonstrates the model's incredible potential. On the HealthBench Professional test, GPT-5.6 Sol scored 60.5 points versus 59 for GPT-5.5, surpassing the average doctor score of 43.7.
My analysis: Concise prompts are a logical step for more powerful models that no longer need excessive hints. However, the incidents with file and subscription deletions are a warning sign. The market must understand that as AI agents become more autonomous, responsibility for safety and proper access configuration shifts to the user. GPT-5.6 Sol is a powerful but dangerous tool, and without a proper safety culture, its use could result in serious losses.