OpenAI Unveils 'GPT-6 Astra,' Enhancing Reasoning and Coding Performance

[Financial News] OpenAI has unveiled its next-generation frontier model, 'GPT-6 Astra,' with enhanced reasoning and coding performance. The model is designed to improve agents' ability to perform complex tasks independently and deliver stronger performance in specialized fields such as science and cybersecurity.
OpenAI said on the 4th that GPT-6 Astra combines advances in pretraining, reinforcement learning, and alignment research to strengthen its problem-solving abilities and its capacity to understand and follow users' intentions.
Astra scored up to 99.9% on the abstract reasoning evaluation ARC-AGI-3 and 97.6% on the advanced mathematics evaluation FrontierMath Tier 4. It scored 57.9% on the coding evaluation Terminal-Bench 4.0 and 74.1% on the software engineering evaluation DeepSWE v1.1.
On OSWorld 2.0, which evaluates computer-use capabilities, Astra scored 72.6%, surpassing GPT-5.6 Sol's 65.7%. In simulations that accounted for latency, the time required to complete tasks was also reduced by about 47%. In science and medicine, it scored 64.6% on Terminal-Bench Science 0.1 and 63.4% on HealthBench Professional.
As its cybersecurity capabilities have improved, OpenAI has also expanded its safety measures. Astra reached the critical level for cybersecurity capabilities under OpenAI's risk management system, the Preparedness Framework, and scored 100% on Exploit Bench, which measures the model's own cybersecurity capabilities. OpenAI also strengthened safeguards, including monitoring systems that detect dangerous cyber activity and halt tasks.
GPT-6 Astra will initially be made available to select organizations. OpenAI plans to expand access to ChatGPT Plus, ChatGPT Pro, ChatGPT Business, and ChatGPT Enterprise users, as well as through the OpenAI API and Amazon Web Services (AWS). The standard API price is $10 per 1 million input tokens and $50 per 1 million output tokens.
[email protected] Choi Hye-rim Reporter