With the launch of the GPT-6 Astra model by OpenAI, the company's president, Greg Brockman, declared that the so-called 'General Artificial Intelligence' (AGI) era had begun. However, the executive also emphasized that there are various definitions regarding what constitutes AGI.
Previously, the most advanced model developed by the creator of ChatGPT was GPT-5.6 Sol, which is now being replaced by GPT-6 Astra. This raises the question of which model is superior and whether the AGI designation represents a real advancement or merely a marketing gimmick.
Fabrício Carraro, Program Manager at Alura and an AI specialist, evaluated the models through his own test. He uses a generic prompt to ask the model to create a space exploration game optimized for graphical quality. Among the tested models were OpenAI's GPT-5.6 Sol and Anthropic's Fable 5, the developer of Claude.
When applying this test to Astra, Carraro noticed a significant difference. He reported that in previous comparisons, Fable 5 had performed better, but with the inclusion of Astra, a great evolution compared to GPT-5.6 Sol could be observed, both in graphic quality and in the choices made by the model.
Furthermore, other users commented online that GPT-6 Astra appears superior to Sol in its ability to achieve a defined goal without giving up, which implies continuously refining the requested details until the desired result is reached.
However, Carraro ponders that the relevance of this advance to daily routine depends on the use; for everyday tasks such as writing reports or generating old images, such superintelligence may not be necessary, given that most daily problems do not require an intelligence level as high as that presented by Sol, Opus, or Fable.
Regarding the definition of AGI, Carraro points out that no one has a technical consensus on the concept. When asked about the AGI label applied to GPT-6 Astra, Brockman stated that 'everyone has a different definition,' although he does not consider it an exaggeration to say that OpenAI created its own version.
Carraro adds that Brockman could argue that GPT-5.2, launched in December 2025, could already be considered AGI, given the absence of a real and technical definition for the term.
Dr. Fei-Fei Li is an open critic of this Silicon Valley obsession with AGI, considering that the discussion is amplified by marketing and corporate enthusiasm. She is recognized as one of the world's most influential computer scientists, a full professor at Stanford University, and co-director of the Institute for Human-Centered AI – HAI.
In terms of functionality, GPT-6 Astra presents a controversy related to its way of 'thinking' or reasoning, involving a technique called 'opaque recurrence' (or looped Transformer).
Traditional AI systems that reason step-by-step generate reasoning tokens, explicitly writing out steps such as: 'First I need to discover X, then I calculate Y to reach answer Z.' This allows experts to audit the correctness and safety of the AI. GPT-6 Astra operates differently: it repeats its information in a cycle within the same internal processing block before generating any text, reasoning internally in the neural network activations without translating these intermediate steps into textual language.
OpenAI acknowledged in its system safety card that GPT-6 Astra presents a 'substantial drop in the monitorability of its reasoning chain' compared to GPT-5.6 Sol. The company confirmed that Astra is significantly better at controlling and hiding parts of its own thought process, making it 'opaque' to external auditors. Despite this, OpenAI positions this model as the safest it has ever developed.
