The 'Danger' Label Attached with the AGI Declaration, The Two Faces of GPT-6 Astra

▲ Photo: AI-generated image

GPT-6 Astra, unveiled by OpenAI on the 3rd (local time), simultaneously earned two contrasting titles. Greg Brockman, OpenAI President, declared upon announcing this model that "we have entered the era of AGI (Artificial General Intelligence)," but the same model received a 'critical' rating, the highest risk level in cybersecurity, for the first time among all OpenAI models.

Astra was developed by integrating pre-training, reinforcement learning, and model alignment research, achieving top-tier benchmark scores in computer usage, web browsing, software engineering, and scientific research.

It scored 98% in FrontierMath Tier 4, a math evaluation metric, 99.9% in ARC-AGI-3, and 100% in ExploitBench, which measures the ability to detect security vulnerabilities. The problem lies with this last metric. It was confirmed that the model could autonomously identify system vulnerabilities and even create attack tools without specific instructions, leading OpenAI to assign it the highest risk rating in the cybersecurity field.

In the initial stages of its release, there was also confusion regarding access. Many paid ChatGPT users could not access the model immediately after service launch, and only companies registered in the Daybreak cybersecurity program were granted priority access. Sam Altman, OpenAI CEO, posted an apology on his social media account for the "confused launch," after which access was gradually expanded. However, a week later, the opposite problem emerged. Demand exceeded expectations, forcing a temporary suspension of new subscriptions for 'ChatGPT Pro,' the top-tier plan costing around $200 per month. OpenAI explained this as a minimum measure to maintain service quality for existing users.

The pricing structure also differed from previous models. Astra's standard fee is $10 per million input tokens and $50 per million output tokens, a 2.5-fold increase compared to the previous model. However, this price includes a 'Fast' mode that offers up to double the speed, and it is assessed that the increased unit cost is partially offset by processing efficiency.

Domestic reactions were far from quiet. In the week following its launch, several reviews appeared in marketer and developer communities, stating that even chatrooms unrelated to work were filled with discussions about Astra. Along with user feedback like "tasks are completed in an instant when using the speed feature," jokes circulated, such as "Kospi is rising because of GPT." OpenAI's official launch post garnered over 36 million views in just two days, making it the most impactful announcement since the unveiling of Sora in OpenAI's history.