en.Wedoany.com Reported - OpenAI released the GPT-5.6 Sol series of models last week. This series adopts a layered architecture, comprising three sub-models: Sol, Terra, and Luna, designed for high reasoning capability, balanced performance and cost, and low cost with high speed, respectively. The new series has achieved improvements in challenging tasks such as programming, biology, and cybersecurity, and is equipped with multi-layered safety protection mechanisms.
For regular users, the ChatGPT platform automatically selects the most suitable model for each conversation; businesses and developers can choose via the API. Sol is suitable for complex problems and deep analysis, being slower but with the highest cost; Terra is suitable for daily advanced tasks, representing a compromise between performance and cost; Luna is ideal for large-scale response generation and simple classification tasks, with the lowest cost. OpenAI stated that after weeks of intensive testing, the model is currently only available to a select group of enterprises, researchers, and strategic clients, and has not yet been opened to the public. This aligns with the US government's trend of restricting very advanced artificial intelligence models before they are fully released, similar to the situation recently encountered with Anthropic's models Mythos and Fable.
GPT-5.6 Sol has improved efficiency on difficult tasks, consuming fewer computational resources and lower costs when solving complex problems compared to its predecessor. The core improvement lies in the reasoning system: the model evaluates steps, tests multiple paths, and optimizes problems before answering, while also executing different tasks in parallel or sequentially through internal modules. This method allows the model to break down tasks into steps, determine the execution order, and invoke system commands, APIs, files, and databases. In scientific tasks, the model has made progress in genomic analysis and quantitative modeling; in the field of cybersecurity, its ability to discover vulnerabilities in systems, code, or configurations has been enhanced, drawing attention from the White House.
OpenAI stated that GPT-5.6 Sol is equipped with multi-layered safety protection mechanisms. In addition to built-in model filters, it also includes real-time monitoring and user behavior analysis. The system intervenes mid-process when detecting problematic prompts and tracks users' historical behavior to identify potential abuse. US authorities are concerned about the strategic impact of these advanced models in cyberattacks and vulnerability exploitation, while OpenAI believes that restricting access could hinder progress in safety and innovation.
The company expects the model to be opened to the public in the coming weeks. The usage cost is calculated per million tokens: Sol data input $5, output $30; Terra data input $2.5, output $15; Luna data input $1, output $6.
