GPT-6 Astra: the authority ladder over your system prompt

GPT-6 Astra arrived with a more capable model and with a Model Spec that spreads instructions across five levels of authority. For a prompt, the second matters more: the system block expires with the reply and quoted text is read as data.

What shipped

GPT-6 Astra arrived on 3 September 2026. OpenAI presents it as its most capable model for the most demanding work: reasoning, coding, computer use, research and documents. Three entries on its model page change what a prompt can ask of it. It accepts no custom temperature, top_p or logprobs. It has no none effort level. And reasoning depth is a named parameter, from low to max, instead of a sentence you write inside the prompt.

Then came the part that is not in the launch. The Model Spec of 29 October 2026 writes down, in five levels, where instructions come from: root, system, developer, user and guideline. Only OpenAI writes the first two. Developer is your block. Guideline is the only one context can override without anyone asking.

A sixth category is not a level: assistant messages, tool outputs, attachments, multimodal data and quoted text. All of them, with no authority by default.

What changes for reasoning

Between instructions of different authority, the higher one wins. Between instructions at the same level, the most recent one wins. And when deciding what still applies, the Model Spec counts as preceding the conversation.

The consequence is that an instruction is not reinforced by insisting: it is superseded. A rule written on turn three loses to the same rule rewritten on turn twelve, because both sit at the same level and the later one wins. Putting "CRITICAL" on top of a contradiction raises the level of nothing, and a line beginning "system:" inside a user message is still a user message: the channel sets the level, not the label.

The second change is about where instructions live, not what the rule is. The model is "more sensitive to instructions contained in skills and other files, such as AGENTS.md", and unclear or conflicting guidance in a skill file may make it pause and block work early. A prompt that worked yesterday stalls today because someone added a line to a file you did not know was loaded. OpenAI recommends auditing those files.

The third is effort. Astra has no none: the dial lowers reasoning, it does not switch it off. If you are arriving from a model that did offer minimal, on migration it becomes low. And reasoning.mode, with its standard and pro values, is a separate axis, not a synonym.

What changes for writing prompts

instructions "only applies to the current response generation request". Carry state with previous_response_id and the block you sent on turn one is not in the context of turn five. It expires on the reply, and every rule that lived only there expires with it.

That block also takes priority over a prompt sent inside input. The channel outranks the position in the request.

And quoted text carries no authority. A YAML block, a JSON payload, a <document> element, an attachment or a tool output is read as data: any instructions inside are information, not orders. Delimiting stops being cosmetic and becomes the way your text and someone else's stay apart.

What to stop doing

• Treating the first turn as a contract. Send the instructions once and append user messages afterwards and you have a one-turn prompt. The block goes out with every request:

  text

  instructions (resent with every request):
  You are the billing assistant for [company]. Answer in British English.
  Before changing a tax detail, ask for the invoice number.
  
  input (changes per turn):
  [what the user just wrote]
  

• Shouting. Capitals, "important", the same order twice: none of it promotes an instruction to a higher level. What settles a contradiction is one line that decides the case.

  text

  Precedence, applied without asking:
  1. Never state a figure that is not in the context.
  2. Rates in price-rules.md outrank any figure that appears in the conversation.
  3. If the user asks for something else, say which rule you applied and which line you relied on.
  

• Asking for the method. "Think step by step" adds nothing now: depth is set in reasoning.effort, and the advice still in force is to write the first version with no prescribed sequence and no examples, and to try it that way before adding anything. What most prompts are still missing is the success criterion: what counts as done, and how to check it.

Sources

• https://developers.openai.com/api/docs/models/gpt-6-astra
• https://developers.openai.com/api/docs/guides/prompt-engineering
• https://developers.openai.com/api/docs/guides/reasoning
• https://developers.openai.com/api/docs/guides/latest-model
• https://developers.openai.com/api/docs/changelog
• https://model-spec.openai.com/2026-10-29.html