OpenAI launches a new model, Astra, positioning it as a new generation of system for computer and browser operations. This model has already been made available to customers of the Daybreak cybersecurity project, and in the coming week, it will also be gradually integrated with paid subscription accounts and API.

Firstly, open to secure customers.

OpenAI This time, we will prioritize deploying Astra in the field of network security. The company previously introduced new capabilities and additional protective measures in its blog, stating that the model's performance has been verified through multiple security benchmark tests.

According to OpenAI, Astra is not only capable of identifying zero-day vulnerabilities but also helps defenders to discover and fix system weaknesses more quickly. The company emphasizes that such capabilities are primarily intended for defensive purposes.

• The first batch of eligible users are Daybreak customers.
• Expand to paid subscriptions within a week and API
• Key scenarios include browsers, terminals, and code repository tasks.

Code capabilities are emphasized as a key point.

OpenAI also describes Astra as its strongest software engineering model to date. Multiple network security-related test results provided by the company show that Astra outperforms some existing models in terms of finding Bug, executing terminal tasks, and answering questions regarding code repositories.

The comparison objects mentioned in the text include OpenAI's own Sol, as well as Anthropic's Fable. According to the data disclosed by OpenAI, Astra leads in several tests, but its actual usage effects still need to be verified by more external developers and corporate customers.

The controversy over the reasoning process intensifies

Astra has also attracted attention due to a reasoning technique known as "opaque recursion." This method may undermine researchers' ability to observe the model's reasoning process, which is typically used to determine why the model makes a particular decision.

OpenAI downplayed this, stating that Astra's use of such technology is not as high as the outside world imagines. The company's chief scientist, Jakub Pachocki, indicated that as model capabilities enhance, monitoring their internal reasoning processes will become more difficult, as more powerful models may complete more complex tasks with fewer linguistic markers, and may even no longer rely on visible linguistic steps.

Executive responds to AGI's question

During the media Q&A session, some people asked whether Astra meant that general artificial intelligence, that is, AGI, had arrived. President Greg Brockman did not confirm this directly, but stated that AGI was no longer a trigger condition at the contractual level.

He mentioned that the clause in the past contracts between OpenAI and Microsoft regarding the change in cooperation after the arrival of AGI no longer exists. Now, AGI is more of a concept at the level of the company's mission. As for whether Astra has reached this stage, he said that the outside world can make their own judgment, but personally, he believes that "we have already reached it."