Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableThe Decoder→ original

OpenAI Revised Its Prompts for GPT-6.2: Fewer Questions, Fewer Tests

OpenAI recommends revising system prompts when moving to GPT-6.2 Astra and Sol. By default, the new models write shorter answers, like to ask permission for obvious steps, and write code sparingly. The company has published ready-made templates in its documentation: act on reversible assumptions, reuse code from the repository, and write tests only on request. If you edit the conversation history in the API, however, the model's earlier reasoning is erased and token usage grows.

AI-processed from The Decoder; edited by Hamidun News
OpenAI Revised Its Prompts for GPT-6.2: Fewer Questions, Fewer Tests
Source: The Decoder. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Decoder ↗
Editor-in-Chief
Key metric #1Models stop only before irreversible actions or a real blocker
Key metric #2Tests are written only on request, when the repository requires them, or to guard against a regression
Key metric #3The temperature and top_p parameters are no longer supported in GPT-6.2

Anyone who has moved working prompts to a new model knows the effect: what worked perfectly yesterday produces strange results today. With GPT-6.2, it happened again. OpenAI explicitly advises developers who are moving to Astra or Sol to re-read their system prompts. To help, the company has published ready-made templates in its documentation that address three habits the new models have shown: answers that are too short, unnecessary questions, and excessive fuss around code.

Some background. Astra is designed for the most demanding tasks, while Sol handles complex work at a lower cost. Both launched on November 17. In the API, Astra starts at $8 per million input tokens and $40 per million output tokens for contexts up to 272 thousand tokens. Beyond that threshold, pricing rises to $16 and $60, and cache reads cost $0.80 per million tokens. For Astra 6.0, the rates above this threshold are higher still: $20 and $75.

Shorter by default: the old "be concise" instruction is no longer needed

Astra writes more briefly and more directly than its predecessors by default. For that reason, instructions like "be concise" should be reconsidered: the model is already concise, and such a request can make the answer too dry. OpenAI offers a template for a working assistant. In it, the model uses the user's and the source's terminology, writes in the active voice, and puts the main point first. For external readers, it must remember that they do not have your context. When editing someone else's text, it must preserve the author's meaning and wording.

A separate problem arises on long tasks. The model sometimes carries working shorthand from its own reasoning, or from messages from other agents, into the final answer. The reader sees notes, not the result. The fix in the template is this: the final answer is the user's first look at the work, and it must read on its own, without intermediate comments that the interface collapses.

Fewer questions: act on assumptions

The second complaint concerns initiative. When the user's intent is unclear, GPT-6.2 sometimes proposes the next step and waits for confirmation. In agentic scenarios this is frustrating: the chain stalls for no reason. OpenAI's recommendation is to let the model carry out such steps itself when it has the authority to do so.

It should stop only in four cases:

  • the scope of the task changes significantly;
  • the model hits a real blocker;
  • a decision is needed that only a human can make;
  • an irreversible or destructive action is about to be taken.

Otherwise, the model should make reversible assumptions, flag them honestly, and keep working. According to the company, a higher reasoning level helps it complete final checks on long tasks.

Code: reuse, don't abstract, tests on request

The programming template contains several rules. First, use the repository's existing classes, methods, and APIs, provided this does not break correctness. Before deleting or merging code, preserve the behavior that calling code depends on. Second, do not write new tests unless asked, unless the repository requires them, or unless they are needed for a specific regression. Third, code must be readable by humans: short variable names and one-liners do not substitute for simplicity. Fourth, do not multiply helper functions and abstraction layers.

There is a noticeable shift in tone. In September, Astra would launch thorough testing over the smallest edits; now it is told not to. For front-end work, developers are advised to specify images and typography explicitly: by default, the template prefers realistic images from generation or search, SVG only on request, and no decorative numbering or badges.

What changed in the API

The most important point for production. Starting with GPT-6.2, the Responses API discards the model's earlier reasoning if you edit the conversation history, switch to some other models, or change the end user or organization. Quality may drop and token usage may rise, because the model repeats work it has already done. For new integrations, OpenAI recommends the "medium" reasoning level, and for those currently on "none" or "minimal," it advises moving to "low." The temperature and top_p parameters no longer exist.

What to do

If you are upgrading, do not copy old prompts blindly. Remove duplicate requests for brevity, add rules about assumptions and stopping points, and check whether you are truncating conversation history. Also watch the 272-thousand-token threshold: long sessions cross it quickly, and the output token price jumps. In such cases, Codex automatically compresses the context. According to Ted Sanders of OpenAI, the choice between Astra and Sol depends on budget and quality requirements. All of these recommendations currently come from OpenAI itself, and there are no independent measurements, so test them on your own tasks.

Story

Key Questions Answered

What are the key details on Shorter by default: the old "be concise" instruction is no longer needed?

Astra writes more briefly and more directly than its predecessors by default. For that reason, instructions like "be concise" should be reconsidered: the model is already concise, and such a request can make the answer too dry. OpenAI offers a template for a working assistant. In it, the model uses the user's and the source's terminology, writes in the active voice, and puts the main point first. For external readers, it must remember that they do not have your context.

What are the key details on Fewer questions: act on assumptions?

The second complaint concerns initiative. When the user's intent is unclear, GPT-6.2 sometimes proposes the next step and waits for confirmation. In agentic scenarios this is frustrating: the chain stalls for no reason. OpenAI's recommendation is to let the model carry out such steps itself when it has the authority to do so.

What are the key details on Code: reuse, don't abstract, tests on request?

The programming template contains several rules. First, use the repository's existing classes, methods, and APIs, provided this does not break correctness. Before deleting or merging code, preserve the behavior that calling code depends on. Second, do not write new tests unless asked, unless the repository requires them, or unless they are needed for a specific regression. Third, code must be readable by humans: short variable names and one-liners do not substitute for simplicity.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…