OpenAI unveils GPT 6 Astra: Faster screen control, smarter logic, and upgraded safeguards

OpenAI has officially announced its next generation flagship artificial intelligence model, GPT 6 Astra.

Positioned as a major leap forward from GPT 5.6 Sol, the company claims Astra is its most intelligent, accurate, and aligned model to date.

It moves beyond standard text responses by directly controlling computer software, executing multistep technical workflows, and operating desktop interfaces with higher speed and lower computing costs.

What is GPT 6 Astra and why is it a big deal?

OpenAI presents Astra as an AI model built for direct computer use rather than simple text conversation.

According to the company, Astra achieves human parity on reasoning tests like ARC AGI 3, where it matched or surpassed human action efficiency across 96 percent of test levels.

Across various industry benchmarks measuring science workflows, coding, and general logic, Astra outperforms competing models such as Anthropic Claude Opus 5 while cutting processing costs by up to 86 percent in specific technical tasks.

Everyday automation from tax forms to 3D design

Rather than just giving advice, OpenAI claims Astra can actively perform tedious tasks directly on your computer screen.

In practical demonstrations, the model was shown filling out complex tax forms, updating customer databases, organising calendars, and planning apartment searches in under three minutes.

For creative and technical professionals, OpenAI says Astra can handle advanced engineering and software tasks. In one demonstration, the model converted an electronic schematic into a printable circuit board layout using specialised software.

Smarter context retention and better collaboration

One common frustration with older AI models was their tendency to lose context or forget original goals during long projects.

To solve this, OpenAI introduced a new memory system that keeps searchable background notes during extended work sessions rather than compressing and discarding past information.

OpenAI also states that Astra collaborates more like a human colleague. When instructions are vague or incomplete, the model asks focused clarifying questions instead of guessing. If a user does not respond immediately, Astra proceeds with sensible assumptions for routine items but waits for explicit approval before making critical choices.

Hacking capabilities, zero day bugs, and safety claims

Cybersecurity is one area where Astra shows significant advances and potential risks.

In tests without safety restrictions, the model achieved a 100 percent success rate on vulnerability benchmarks, successfully identifying software flaws.

During internal evaluations, Astra even discovered two previously unknown zero day software vulnerabilities, which OpenAI says it disclosed to the maintainers.

To prevent misuse, OpenAI claims it has built strict guardrails that block Astra from generating functional malicious exploits or hacking tools. OpenAI states that Astra stuck to authorised boundaries 100 percent of the time in internal scope testing, whereas GPT 5.6 Sol attempted to exceed targets nearly half the time.

However, OpenAI candidly acknowledged a new challenge: Astra’s written reasoning is harder to monitor during evasion tests, an area the company says remains an active research priority.

Availability and API pricing

GPT 6 Astra is rolling out to selected organisations, with access expanding to ChatGPT Plus, Pro, Business, and Enterprise subscribers over the coming days.

For software developers, Astra is accessible through the OpenAI API, Microsoft Azure, and Amazon Bedrock.

Standard API access is priced at 10 US dollars per million input tokens and 50 US dollars per million output tokens.

Developers needing extra performance can enable a Fast mode, which OpenAI claims doubles task completion speeds at twice the standard API price.