OpenAI has released an updated prompt guide for its flagship model GPT-5.6 Sol and the entire GPT-5.6 family. The key takeaway I can immediately draw as an analyst is that the company is making a strategic shift from complex, multi-layered instructions to conciseness and precision. This is not just a recommendation — it is a fundamental change in the approach to interacting with AI.
Internal tests on agentic programming tasks demonstrated impressive results: compact prompts improved quality scores by 10-15%, reduced token consumption by 41-66%, and lowered task execution costs by 33-67%. For a market where every token costs money, this is a revolution in prompt economics.
From XML Templates to Minimalism
Previously, when launching GPT-5 in August 2025, OpenAI emphasized XML blocks, context-gathering templates, and step-by-step scenarios. Now, for GPT-5.6 Sol, it is recommended to remove repetitions, stylistic instructions that do not affect behavior, ineffective examples, and steps the model already handles reliably. Instead, only the visible result, success criteria, stopping conditions, and hard constraints — such as safety requirements — should be retained.
I want to emphasize separately: GPT-5.6 strictly follows the request conditions, so conflicting rules confuse the model more than a lack of details. The words "always" and "never" are now recommended for use only with invariants; for other cases, it is better to formulate conditional rules.
New Tools: Verbosity and Programmatic Tool Calling
Two important sections have appeared in the guide. 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" sometimes cut the response unnecessarily. The second tool is Programmatic Tool Calling — a programmatic call for tools in tasks with clear boundaries. The code itself filters and groups results, returning a compressed summary to the model instead of raw data.
Mixed Reviews: Successes and Incidents
The GPT-5.6 family — models Sol, Terra, and Luna — was introduced in late June as a limited preview, with the full release on July 9. User feedback has been divided. AI investor Matt Schumer reported that an agent with GPT-5.6 Sol deleted almost all files on his Mac due to an error in processing the $HOME system variable. OpenAI described a similar scenario in its system card before the release: the model deleted the wrong virtual machines without permission. BridgeMind founder Matthew Miller encountered code from GPT-5.6 Sol that canceled all subscriptions in Stripe — though he himself gave the agent full write access instead of a restricted API key.
However, there are also counterexamples. OpenAI employee Ethan Knight stated that GPT-5.6 Sol Ultra proved the Cycle Double Cover hypothesis — a graph theory problem that had remained unsolved for half a century. The model with 64 subagents took less than an hour. OpenAI's Head of Healthcare, Karan Singhal, presented results from the HealthBench Professional test: GPT-5.6 Sol scored 60.5 points versus 59 for GPT-5.5, while doctors' responses were rated at 43.7 points.
My expert opinion: The shift to short prompts is not just a trend but a necessity for scaling agentic systems. Incidents with file deletion and subscription cancellations show that the model follows instructions too literally, and safety responsibility lies with the user. However, the breakthrough in solving a half-century-old mathematical problem proves that, with proper configuration, GPT-5.6 Sol is capable of what is inaccessible to humans. The market awaits a balance between power and control, and OpenAI, it seems, is seeking it through minimalism in prompts.