OpenAI has unveiled GPT-6 Astra, a next-generation frontier model that elevates performance across all domains including reasoning, coding, science, and cybersecurity. OpenAI explained that it is a model that has simultaneously strengthened complex problem-solving capabilities and user intent understanding ability by integrating outcomes from pretraining, reinforcement learning, and alignment research.

In the coding domain, it achieved Terminal-Bench 4.0 57.9% and software engineering evaluation DeepSWE v1.1 74.1%.

Beyond benchmarks, OpenAI stated that the model contributed to solving long-standing unsolved problems in actual mathematical research.

It recorded 72.6% on the OSWorld 2.0 offline evaluation, exceeding GPT-5.6 Sol's 65.7%, and when latency was reflected in simulation, task completion time was reduced by approximately 47%.

On AutomationBench, the complex task workflow evaluation, it showed 41.4%, more than double GPT-5.6 Sol's 18.1%.

On both GPQA Diamond and HealthBench Professional, OpenAI stated it achieved the highest scores among models listed in the public comparison table.

On ExploitBench, which measures model capabilities alone by excluding operational environment safeguards, it recorded 100%.

OpenAI stated that while this capability can be utilized for discovering and fixing security vulnerabilities, it also carries risks of misuse.

Accordingly, it explained that it has expanded monitoring and protection measures for detecting, suspending, and halting dangerous cyber activities and for deterring model misuse.

Standard API pricing is $10 per 1 million input tokens and $50 per 1 million output tokens.