OpenAI tells developers to change prompts for GPT-6.2 Astra and Sol
The models give short answers and put less questions. A change to chat history in the API erases reasoning.
OpenAI tells developers who move to GPT-6.2 Astra or Sol to examine their system prompts. A system prompt is the text that sets how a model works. The company put prompts that you can copy in the documentation. OpenAI launched the 2 models on November 17.
How it works
- Astra writes short answers by default. An instruction such as be concise can cut too much.
- A new prompt tells the model to do the next steps that are approved, and to make assumptions.
- The code prompt tells the model to use the code that the user has. It writes new tests only when the user tells it to.
The numbers
OpenAI gave no test results for the prompts. The API changed. The Responses API erases the reasoning that the model did before. This occurs if you change the history, or move to some other models. This can decrease quality and increase token use. OpenAI recommends medium reasoning for new work.
What is not known
OpenAI gave no data on how much the prompts help. It gave no data on erased reasoning. Personnel must test this on their work.
Insight: Each OpenAI release makes developers write prompts again
In April, OpenAI told developers to start with new prompts for GPT-5.5. In September, it gave GPT-6 Astra prompts to stop follow-up questions. It gives the same fix again at this time. On November 2, users saw GPT-6.2 Astra in Codex with no notice. This reads as work that moves to builders. The vendor changes defaults, and you do the prompt work. Keep a set of tests and record the model name.
Sources
Posted