GPT-5.6 Tells Your Bloated System Prompt to Sit Down and Shut Up 🪶

OpenAI's newly published GPT-5.6 Sol prompting guide pushes developers toward outcome-first prompts, telling them to drop the multi-page scaffolds that became standard with the August 2025 GPT-5 launch. The core instruction: define what "done" looks like, set stopping conditions, and let the model handle the rest.

Internal coding-agent benchmarks back the change. Trimming redundant instructions and repeated style rules raised evaluation scores by roughly 10–15%, cut total tokens by 41–66% and reduced costs by 33–67%. XML persistence blocks, context-gathering templates and tool-preamble scripts that once helped coordinate reasoning are now treated as noise the model has to parse around.

The guide's reference prompt begins with the user-visible outcome — "Resolve the customer's issue end to end" — followed by explicit success criteria, required pre-response actions and fallback behavior when evidence is missing. The model is expected to manage process steps on its own, with developers supplying only destination, constraints and stopping rules.

OpenAI also flags a new risk: GPT-5.6 follows prompt contracts closely, and "conflicting rules can create more instability than missing detail." Where earlier models simply picked one rule when instructions clashed, GPT-5.6 burns reasoning tokens trying to reconcile both, slowing output and increasing error rates. The guide singles out overlapping "always" and "never" absolutes as a common source of that instability.

Two new tooling features round out the update. The text.verbosity parameter replaces older "be brief" instructions, which now over-correct because GPT-5.6 is more concise by default than GPT-5.5; developers can set a global default and override per task. A Programmatic Tool Calling section covers bounded workflows in which code handles filtering, batching or aggregation before the model is invoked.