GPT-6 Astra: What’s Actually New in OpenAI’s New Frontier Model

OpenAI has introduced GPT-6 Astra, their latest frontier model, just a week after Anthropic unveiled Claude Fable 5.1. The company touts Astra as the most intelligent and aligned model ever created, emphasizing that it goes beyond merely answering questions – it can carry out tasks independently. Instead of merely suggesting how to complete a task, Astra can complete it itself, generate complete documents without a rough draft, retain context over extended coding sessions, and make judicious decisions about when to act and when to stop.

One significant enhancement is Astra's ability to interact with computers. It can fill out forms, update CRM records, run frontend quality assurance checks on websites, and troubleshoot software by observing screen activity, all without needing explicit instructions for each step. During testing, Astra scored 72.6% on OSWorld 2.0, surpassing Claude Opus 5's 70.2% and outperforming GPT-5.6 Sol's 65.7%.

Another improvement is Astra's ability to discern when to ask questions and when to provide answers. Unlike earlier models that either guessed incorrectly or asked more questions than necessary, Astra is trained to fill in routine gaps on its own, pausing only when the response could significantly change the outcome. For example, in a side-by-side demonstration, Astra paused after 20 seconds to ask what career a user was entering into a personal career website builder, showing the difference between a model that takes action and one that uses judgment.

Astra also excels at producing finished documents that adhere to the user's templates, tone, and structure while using only relevant context, rather than adding extraneous details. It can create a slide deck from a few template slides while maintaining the consistent tone and layout throughout. Additionally, in long coding sessions, Astra can maintain searchable notes across context windows, preserving important details that might otherwise be lost due to the model's compression of lengthy debugging sessions into a single summary.

Perhaps most notably, Astra has achieved a "Critical" level of cybersecurity readiness according to OpenAI's Preparedness Framework. This means the model can independently identify and develop working exploits for previously unknown vulnerabilities. As a result, OpenAI has set a higher access threshold for Astra. While it will assist with defensive tasks like secure code review and patch validation, it will not create proof-of-concept exploits at launch.

Instead, access to this capability will be restricted to those who participate in OpenAI's Daybreak program.

Benchmark-wise, Astra leads in computer use tasks, outperforming Claude Opus 5 on OSWorld 2.0 and showing substantial improvements in coding scores over its predecessor. However, its lead over Claude Fable 5.1 on Terminal-Bench 4.0 is narrow. In the Humanity's Last Exam with tools benchmark, Astra scores 57.2%, placing it behind both Claude Fable 5.1 (65.0%) and its own successor, Claude Opus 5 (63.6%).

Some benchmark scores are based on evaluation setups that don't fully represent typical usage, such as GPT-6 Astra's 99.9% on ARC-AGI-3, which requires an expensive, stateful evaluation harness. Independent testing by the ARC Prize Foundation revealed that a standard, stateless API call scores much lower, between 17% and 63% depending on the reasoning tier used.