From GPT to AGI? Meet GPT-6 Astra: The Model That Starts Working in Your Place

OpenAI launches its most powerful model yet, with autonomous capabilities that change the rules of the game in computer control, cyber, and solving complex problems.

For years we talked about artificial intelligence as a tool that helps us write or think. Today, with the official launch of GPT-6 Astra, we are officially moving into the era of autonomous agents. It is no longer just a chatbot sitting inside a browser window. It is a model designed from the ground up to "hold the mouse" and carry out tasks from end to end inside our operating systems.

The biggest news in Astra is its ability to act as an independent agent. OpenAI put the emphasis on performing actions that until now required a lot of human intervention: filling out complex forms, working inside CRM systems, managing calendars, deep web research, and even running QA tests on user interfaces.

According to VentureBeat, in the OSWorld 2.0 test, the new model completed tasks in 47% less time compared with its predecessor, GPT-5.6 Sol. While Sol needed about 75 minutes for a complex task, Astra finishes it in only 40 minutes with a significantly higher success rate.

For developers and data people, Astra brings critical improvements to the Codex model. Its new ability to save and search Context from long sessions, without compressing the entire history into one summary that loses details, is a significant change when working on large projects.

On the enterprise side, the model works better with documents, presentations, and spreadsheets while keeping existing templates and understanding the organization's unique style. It does not just produce text, it understands the business context in which it operates.

The numbers Astra shows in benchmarks are almost unbelievable:

• •99.9% on the ARC-AGI-3 test: A test considered one of the most reliable measures of generalization ability and general intelligence.
• •97.6% on FrontierMath Tier 4: Solving math problems at the level of researchers.
• •100% on ExploitBench: Perfect ability in identifying and exploiting security vulnerabilities.

In fact, Astra is OpenAI's first model to receive a "Critical" rating under the company's Preparedness Framework. During testing, the model managed to find two Zero-Day vulnerabilities that were not known before. Because of this power, OpenAI chose to limit some of its advanced cyber capabilities at the initial launch stage to prevent misuse.

An equally interesting point is the model's level of compliance (Alignment). In tests designed to check whether the model would try to exceed the permissions it was given in order to complete a task, GPT-5.6 Sol exceeded them in 48% of cases. Astra, by contrast, showed 0% overreach. This is a critical figure for organizations that are afraid to give artificial intelligence access to their core systems.

The launch starts today for a limited group of organizations (the Daybreak program), and in the coming days Astra will reach Plus, Pro, Business, and Enterprise users. It will be available both through the API and on Amazon Bedrock.

API costs:

• •Input: $10 per million tokens.
• •Output: $50 per million tokens.

High-end users (Pro and above) will get access to the Astra Pro version, which offers even better performance for especially complex tasks.

Bottom line: We are at a turning point. GPT-6 Astra is no longer just a writing aid, but an autonomous digital workforce. The question is no longer what it can write for you, but which entire processes you are ready to hand over to its responsibility.