OpenAI is set to launch its new model, "Astra," generating significant attention and controversy in the AI research community—it may represent the largest performance leap since GPT-4, but its underlying novel reasoning technology has raised concerns about diminished capabilities for AI safety monitoring.Author and source: The Information
According to tech media The Information on September 2, Astra incorporates an innovative technology that pushes beyond the model’s previous limits in reasoning through complex problems. Some researchers believe Astra’s improvements in coding and computer usage could rival the performance leap introduced by OpenAI’s release of GPT-4 in 2023, far surpassing the performance seen with last summer’s GPT-5 release. This represents a positive signal for the entire industry, which relies on AI-driven productivity gains to justify large-scale investments in data center infrastructure.
However, this technological breakthrough also brings significant risks. The new technology allows models to perform deep reasoning without fully outputting their chain of thought, meaning researchers may no longer be able to effectively monitor the model’s reasoning process as before, thereby weakening existing safety review mechanisms. Currently, OpenAI is taking steps to ensure Astra’s chain of thought remains visible, but the industry remains divided on the long-term implications of this trade-off.
Performance breakthrough: The new inference technique is closely related to Astra’s core technological advancements, including concepts such as "recurrent depth" and "loop transformers."
According to The Information, the technology processes problems by repeatedly passing them through the model’s mathematical “layers” to generate the next word in the answer, enabling the model to demonstrate reasoning capabilities far beyond its own size—equivalent to those of a much larger model—without significantly increasing its parameter count.
Some researchers believe that Astra's performance improvement, driven by its enhanced coding and computer usage capabilities, could be comparable to the groundbreaking leap seen with the release of GPT-4 in 2023, far exceeding the relatively muted market response to GPT-5's launch last summer.
This advancement holds significant importance for the entire AI industry chain. The large-scale construction of data centers worldwide is fundamentally based on the expectation that continuous improvements in AI capabilities will drive productivity growth and translate into higher business returns. If Astra’s performance meets expectations, it will provide strong support for this investment thesis.
It is worth noting that this technology is not a new concept. AI research pioneer Jürgen Schmidhuber pointed out on the social platform X that the core idea of "recurrent thinking" was already presented in his 2015 paper, "On Learning to Think" (arXiv:1511.09249).
He stated that the control network C in the paper is essentially a prompt engineer, learning to query independent neural world models to perform abstract reasoning; the generated prompts and responses are internally self-generated vector sequences that do not need to be expressed in natural language.
Security risk: The chain-of-thought monitoring mechanism faces challenges as new technologies, while enhancing performance, may potentially disrupt existing AI security monitoring frameworks.
Reports indicate that current mainstream AI models typically output their reasoning processes in text form—known as "chain of thought"—when handling complex problems. This mechanism not only enhances model interpretability but also serves as a critical tool for researchers to monitor model behavior and prevent anomalous operations.
According to The Information, chain-of-thought monitoring was one of the primary solutions OpenAI proposed after the July Hugging Face hacking incident to prevent similar security breaches from recurring.
However, the new reasoning technique adopted by Astra may not fully output reasoning steps when the model deeply "thinks," instead completing computations internally in a "silent" manner. This means that the more the model relies on this new technology, the less transparent its reasoning process becomes to external observers.
To address this, OpenAI is currently taking steps to strike a balance. The company is guiding the model to reduce the number of loops at the computational level, ensuring that Astra’s chain of thought remains somewhat visible—the fewer the loops, the less room the model has for “silent thinking.”
Major labs are already following up; the long-term monitoring solution for this trend’s impact may extend far beyond OpenAI alone. According to The Information, this new inference technique has become a hot topic in internal discussions at major AI labs, and the likelihood that organizations such as Anthropic and Google are adopting similar approaches cannot be underestimated.
Several researchers have pointed out that chain-of-thought monitoring was never intended to be the ultimate solution for AI behavior monitoring. As model capabilities continue to evolve, the ability of models to output their reasoning processes in text form is largely a byproduct of current training methods.
As model developers explore new optimization directions and architectural designs, the model's tendency to spontaneously generate chain-of-thought outputs may gradually diminish, requiring researchers to develop new monitoring methods.
The report highlights that the central issue today is whether major AI developers, under increasing pressure to compete on performance, will voluntarily abandon existing security measures in pursuit of stronger reasoning capabilities. The answer to this question will largely determine the direction of the next phase of AI safety governance.