Explanatory journalism with depth and rigorPTENES
Explosão SolarContext. Not just headlines.Search

OpenAI launches Astra and reignites debate on artificial general intelligence

New GPT-6 model delivers advanced benchmark performance and promises to be more autonomous

Daniele Morais
September 5, 2026 · 2 min read
ShareWhatsAppXFacebook
OpenAI launches Astra and reignites debate on artificial general intelligence
Photo: "Artificial Intelligence Talks - WTSA-16" by ITU Pictures is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

Last Thursday, OpenAI announced GPT-6 Astra, its most advanced model to date. The release brought record numbers in artificial general intelligence (AGI) tests and sparked discussions about AI's actual ability to act autonomously.

Why Astra is trending right now

The buzz arose because Astra achieved 99.9% on the ARC-AGI-3 benchmark, designed to measure a system's ability to handle novel problems. This score was cited by OpenAI President Greg Brockman as a milestone toward the so-called AGI era.

Performance in intelligence and mathematics tests

In addition to ARC-AGI-3, the model scored 98% on FrontierMath Tier 4 and 97.6% on the same benchmark, according to the company itself. In broader evaluations, such as Artificial Analysis's Intelligence Index, Astra scored 61 points, virtually equal to GPT-5.6 Sol, indicating that the performance leap is still concentrated in specific areas.

Reduction in hallucinations and greater accuracy

In tests by Artificial Analysis, the hallucination rate – invented responses – dropped from 92% in GPT-5.6 Sol to 51% in Astra. The same analysis pointed to a four-point improvement in the accuracy metric, although there is still room for error in extreme situations.

Autonomous execution capabilities

Astra can operate a computer and browser, create documents, spreadsheets, presentations, and develop websites, completing tasks in up to 47% less time than GPT-5.6 Sol in OSWorld 2.0 simulations. On the same benchmark, its performance rose to 72,6%, compared to 65,7% for its predecessor, reducing execution time from 75 to 40 minutes.

Cybersecurity risks and restricted access

Classified at the "critical" level of cybersecurity capability, Astra was able to identify two unknown zero-day vulnerabilities in internal testing. For this reason, sensitive features will have limited access, and the model will initially be available only to Plus, Pro, Business, and Enterprise subscribers, and via API on Azure and AWS platforms.

What the launch indicates for the market

According to market analyses, OpenAI aims to be the primary AI platform for users, similar to the concept of "primary status" in banking, where customers concentrate their interactions with a single provider. Astra, by combining more accurate responses with practical task execution, reinforces this commercial strategy.

With information from InfoMoney, GZH, VEJA.

Source: InfoMoney, GZH, VEJA

#OpenAI#AI#AGI#GPT-6#cybersecurity#benchmark
Also inPortuguêsEspañol
ShareWhatsAppXFacebook