OpenAI introduced GPT-6 Astra. It absolutely shines in tests, according to the company's president, we have entered the era of AGI

American OpenAI on Thursday introduced its most powerful model to date, GPT-6 Astra. It describes it as a generational leap in capabilities, and its president, Greg Brockman, makes no secret of the fact that, according to him, we are now entering the era of AGI – that is, human-level artificial intelligence.

Best for Computer Control and Programming

OpenAI places Astra at the forefront in several disciplines. According to the company, the model is designed to be best for computer control – it can fill out forms, manage database records, browse the web, or prepare materials in emails and documents. At the same time, it is reportedly the most powerful programming tool, capable of handling complex tasks in real-world projects.

The programming tool Codex, along with Astra, now gets a new way to maintain context during long sessions. Previously, when memory was full, models had to “compress” older work into a summary – but this often lost details about why a fix failed or how a particular piece of code behaved. Instead, Astra keeps ongoing notes across sessions, and older context remains searchable, allowing it to refer back to requests or test results from earlier, even if they weren’t explicitly noted. This is currently an experimental feature that you will need to activate manually in Codex, but it will become default for Astra in the coming weeks.

The company also relies on a series of tests. According to them, Astra satisfied the demanding mathematical benchmark FrontierMath and the abstract reasoning test ARC-AGI-3, where it approached human level.

Astra is launching today. First, selected companies will receive it, then in the coming days, all paying ChatGPT users (Plus, Pro, Business, and Enterprise) as well as developers via API, Microsoft Azure, and Amazon Bedrock. Subscribers to higher tiers will also receive a more powerful version, GPT-6 Astra Pro.

GPT-6 Astra Model Accompanied by Controversies

Astra is the very first OpenAI model to cross the “critical” threshold of cyber capabilities according to the company’s internal security framework. In practice, this means it can independently find and prepare exploits for even previously unknown vulnerabilities (so-called zero-day). This can help defenders patch security holes faster, but it also benefits attackers. The company is therefore releasing Astra with stricter safeguards – partly in response to a recent incident where one of its models went out of control and breached the Hugging Face platform. OpenAI emphasizes that Astra was not involved in that incident.

Questions also hang over how Astra “thinks.” According to available reports, the model uses a technique called opaque recurrence, which makes it difficult to track its internal reasoning (so-called chain of thought) – the process by which researchers check whether the model is acting against their intent. OpenAI downplays the extent of this obfuscation, referring to it as a natural consequence of growing capabilities.

And then there’s the big word: AGI. OpenAI President Greg Brockman stated at a press conference that “we are now in the era of AGI”, meaning human-level artificial intelligence. However, this is his personal assessment, and he himself adds that the conclusion should be left to each reader. It is also noteworthy that, according to the company, its previous models significantly contributed to Astra’s training for the first time.

Do you believe that the era of AGI truly begins with the GPT-6 Astra model?

Sources: OpenAI, The Verge, TechCrunch