"Strengths in Advanced Reasoning and Mathematics" GPT-6 Astra Performance Reviewed... Price Also Up 2.5 Times

OpenAI Unveils GPT-6 Astra

Aiming for Enterprise AI Leadership with Coding and Computer Proficiency

Cost-Efficiency Remains a Challenge

OpenAI has released its next-generation top-tier (frontier) model, GPT-6 Astra. While the company is solidifying its leadership in the enterprise AI market by emphasizing coding capabilities and computer proficiency, some analyses indicate that the increase in token pricing has led to reduced cost-efficiency.

On September 4, OpenAI unveiled GPT-6 Astra, introducing it as a "frontier model that delivers state-of-the-art performance across a range of fields, including reasoning, computer usage, software development, cybersecurity, and science."

Astra demonstrated notable strengths in advanced reasoning and mathematics. In the Abstract Reasoning Challenge (ARC-AGI-3), it achieved a score of up to 99.9%, significantly surpassing its predecessor, GPT-5.6 Sol, which scored 7.8%. In high-level math evaluation (FrontierMath Tier 4), Astra attained 97.6%, outperforming the previous model's 83%.

The model also showed improved performance in coding. In Terminal Bench 4.0, an evaluation of multi-step development tasks, it achieved a score of 57.9%. In DeepSWE v1.1, a benchmark assessing real-world software engineering skills, it recorded 74.1%.

The core advancement of this model is its ability to go beyond simply generating responses—AI now directly interacts with screens and external tools, continuously managing multi-step work. In the benchmark OSWorld 2.0, which assesses the ability to operate an operating system like a human, Astra scored 72.6%, surpassing GPT-5.6 Sol’s 65.7%. In latency-reflective simulations, the time required to complete tasks was reduced by approximately 47%.

Astra is being rolled out first to select organizations, and in the coming days, availability will expand to ChatGPT Plus, Pro, Business, and Enterprise users, as well as being offered through the OpenAI API and AWS.

Alongside performance improvements, the price has also gone up. Astra's standard API pricing is 10 dollars per 1 million input tokens and 50 dollars per 1 million output tokens, which is 2.5 times higher than its predecessor, GPT-5.6 Sol. This suggests that Astra is positioned as a work-focused model intended for jobs requiring complex reasoning and prolonged task execution.

Greg Brockman, President of OpenAI, remarked, "If we look back a few years from now and think about when artificial general intelligence (AGI) was achieved, this moment and this model may well be it."

However, independent research firm Artificial Analysis pointed out that while Astra has improved coding agent performance and token efficiency, it showed little difference from its predecessor in overall intelligence metrics, and, due to the API price hike, general work cost-efficiency has actually decreased.

Major foreign media outlets analyzed that OpenAI is emphasizing software development and professional work capabilities to counter Anthropic, which has increased its presence in the enterprise AI market. The Financial Times (FT) in the UK evaluated OpenAI's launch as an attempt to reclaim the technological lead previously conceded to Anthropic and to reaffirm its position as the market leader in enterprise AI ahead of its IPO.

This content was produced with the assistance of AI translation services.