OpenAI published new prompting guidance for its GPT-5.6 model family on July 9, and the core message reverses a year of accepted practice: stop over-prompting. Instead of the long, rule-heavy system prompts and step-by-step choreography that developers built for earlier models, OpenAI now tells them to state the desired outcome, the constraints, the available evidence and the completion bar, then leave the model room to choose its own efficient path.
The guidance is aimed at developers building on the API and teams running automated agents, and applies across the GPT-5.6 lineup, including the flagship Sol, mid-range Terra and low-cost Luna.
OpenAI backs the shift with numbers, though it frames them carefully. In a sample of internal coding-agent evaluations, it says configurations with leaner system prompts improved scores by roughly 10-15% while reducing total tokens by 41-66% and cost by 33-67%.
The company stresses these ranges are directional and will vary by workload, urging developers to validate on their own tasks rather than treat the figures as guarantees. Still, the direction is notable because it points to a rare case where cutting instructions improves quality and lowers cost at the same time, rather than trading one against the other.
The specific advice is concrete. OpenAI tells developers to state each instruction exactly once and remove repeated rules, redundant examples, inert style notes and process steps the model already handles reliably. It warns against absolute directives like “always” and “never,” reserving them for true invariants such as safety limits, and cautions that GPT-5-class models follow prompt contracts so literally that conflicting rules cause more instability than missing detail.
Rather than prescribing every step, developers should describe what “done” looks like and add explicit stopping conditions. The guide also covers newer features absent from the older playbook, including a text.verbosity setting to control response length and Programmatic Tool Calling for bounded data-processing stages.
Why the Advice Keeps Reversing
The guidance marks the third turn in OpenAI’s prompting philosophy in under a year, and the pattern is revealing. The GPT-5 guide from August 2025 told developers to add scaffolding, with XML persistence blocks, parallel-search templates and verbose tool preambles designed to calibrate how hard the model worked. The GPT-5.5 guide in April already told teams to rebuild rather than port prompts forward.
Now GPT-5.6 asks for even less. Each cycle has pushed developers to write less and trust the model with more of the route, a sign that capability is migrating from the prompt into the model itself, where planning that once had to be spelled out by hand is now built in.
What It Means in Practice
The practical implication is that prompts optimized for older models can now actively hurt results, so OpenAI advises auditing old prompt stacks rather than copying them across, and changing one thing at a time so regressions can be traced.
For enterprises running agents at scale, the token and cost reductions are the headline, arriving just as the industry scrutinizes whether AI spending pays off. There are trade-offs worth noting: leaner, planning-heavy prompts can make the model slower as it maps a problem before acting, and independent developer Simon Willison observed that Sol felt competent without clearly beating Anthropic’s Claude Fable 5 on the complex coding tasks he ran.
The advice is also self-serving in a benign way, since guidance that cuts token use lowers customers’ bills but also eases load on OpenAI’s strained infrastructure. For developers, the takeaway is straightforward: shorter, clearer prompts are now both cheaper and better, provided the changes are validated on real work.
