For years, improving an AI application often meant adding more instructions: another rule, another example, another paragraph explaining exactly what the model should do.
OpenAI’s newest GPT-6 guidance suggests that habit may now be working against developers.
In its practical guide published October 2, OpenAI says increasingly capable models understand nuance and ambiguity well enough that overly specific instructions can hinder results. The challenge is shifting from telling the model every step to designing the environment in which it can make good decisions on its own.
That begins before the prompt. OpenAI now presents GPT-6 as a workload decision: GPT-6 Astra for the hardest reasoning, GPT-6.1 Sol for complex coding, research and computer-use workflows, and GPT-6 Luna for focused, high-volume tasks such as extraction and classification. Developers can then tune reasoning from low through higher effort levels rather than spending maximum compute on every request.
The same thinking extends to cost. OpenAI recommends removing irrelevant context, parallelising independent work, using prompt caching for recurring information and applying compaction when conversations become long. Depending on the model, the company says cached input tokens can cost up to 95% less than uncached input, making context architecture part of the product economics rather than an afterthought.
Prompting changes too. Builders are encouraged to define the outcome, relevant context, constraints, approval boundaries and what completion actually means. Skills and repository instructions such as AGENTS.md should be trimmed rather than accumulated indefinitely, because GPT-6 Astra follows instructions closely enough that old or conflicting rules can become a source of failure.
The guide becomes more consequential when work lasts beyond a single response. GPT-6 workflows can use mid-turn steering to receive updated instructions while running, asynchronous tools to continue independent work during slow operations, and computer use when a task requires interacting with a website or desktop application rather than an API.
OpenAI highlights Harvey, Cognition, Hex and Invideo as examples of companies already applying GPT-6 Astra to legal drafting, software testing, business analysis and video editing.
The message underneath the guide is larger than prompt engineering.
GPT-6 makes asking the AI what to do easier. Building the system that decides what it may do, how far it should go and whether the result was worth the cost is becoming the harder job.
