OpenAI has introduced GPT-6 Astra and calls it the most capable and most aligned model it has released. Alignment means the model is trained to follow the user’s intent, stay inside the assigned task, and avoid unsafe or unauthorized steps. Astra combines pre-training, reinforcement learning, later training that rewards better answers and actions, and extra alignment work. OpenAI says Astra now leads on tests of computer use, meaning it can operate a screen, keyboard, and mouse as a person would, as well as on web browsing, software engineering, cybersecurity, science, and ordinary office work.
The company reports near-perfect scores on several hard tests. FrontierMath Tier 4, ARC-AGI-3, ExploitBench. OpenAI also says Astra helped with unpublished math on the gaps between prime numbers.
Computer use, coding, and safety
The company presents Astra as a system that can fill forms, update records, research a topic, draft email, analyze data, build a simple website, and check that the site works. The announcement presents test results on Agents’ Last Exam, BenchCAD, Terminal-Bench 4.0, GPQA Diamond.
On cybersecurity, OpenAI says Astra has reached the Critical level in its Preparedness Framework, an internal scale for dangerous skills. That means, with tools and access, it can find previously unknown flaws and build ways to use them on well-protected systems without a person directing each step. The public version is said to help with safer tasks such as code review, while refusing more advanced attack work except through a limited defender program called Daybreak. The announcement also says Astra is harder to jailbreak, meaning it is harder to talk into ignoring its safety rules, and less likely to claim skills it does not have.
Astra is opening first to a limited group of organizations. In the following days it is due in ChatGPT Plus, Pro, Business, and Enterprise, and through the OpenAI API, Microsoft Azure, and AWS Bedrock under the name gpt-6-astra. Prices are 10 dollars per million input tokens and 50 dollars per million output tokens.