🛠 OpenAI releases guide for choosing GPT-6 models
In the guide “A model guide for the GPT-6 family,” model selection is described as a balance of “intelligence/price”: Astra — complex reasoning, Sol — coding and computer use, Luna — high-volume repetitive tasks. Reasoning effort in the API is configurable from low to extra high/max and can be changed mid-dialogue without resetting the cache.
🌍 OpenAI shifts the focus from “what the models can do” to “how to build a product”: model selection, input token caching up to 95% cheaper, context compaction, and monitoring are combined into a single process — the path from demo to production.
👤 You can try it right away: reasoning effort can be changed without losing the cache, and computer use works with websites and applications even without an API. These are OpenAI's claims — the guide contains no independent benchmarks.
Source 1: https://openai.com/index/practical-guide-building-gpt-6
Source 2: http://web.archive.org/web/20261002171003/https://openai.com/index/practical-guide-building-gpt-6