OpenAI introduced GPT-6 Astra — a new model for complex professional tasks

OpenAI has introduced its new flagship AI model, GPT-6 Astra. The company describes it as the smartest and most coherent model in its history — Astra demonstrates the best performance among OpenAI models in computer and browser interaction, code writing, and solving scientific and professional tasks. The model scored 98% on the FrontierMath Tier 4 test and 99.9% on ARC-AGI-3.

GPT-6 Astra has already begun deployment in a limited number of organizations, and within the next few days will become available to all users of ChatGPT Plus, Pro, Business, and Enterprise plans, as well as via the OpenAI API and Amazon Bedrock. The standard API pricing is $10 per million input tokens and $50 per million output tokens.

OpenAI specifically highlights Astra’s capabilities in cybersecurity — the model has reached a critical threshold on the risk scale within its internal Preparedness Framework assessment. On the ExploitBench test, the model achieved 100%, compared to 78.5% for the previous GPT-5.6 Sol model. During testing, Astra independently discovered two previously unknown zero-day vulnerabilities — OpenAI has already shared this information with the developers of the affected software. The company states that such capabilities can help security professionals identify and mitigate threats more quickly, but they also require stricter protective measures. At launch, the model refuses to perform advanced tasks such as creating working exploits, but over the coming weeks, OpenAI plans to open broader access to these features for security specialists through its Daybreak program.

Astra is also named OpenAI’s best model for software development — according to the company, it outperforms GPT-5.6 Sol in most internal code-related tests.

However, OpenAI notes a controversial feature of the model — in tests specifically designed to check whether a model can hide its reasoning process from observers, Astra’s chain of thought proved more difficult to track than that of GPT-5.6 Sol. The company considers improving the transparency of the model’s reasoning a priority for future research.

Source: OpenAI