OpenAI Releases GPT-5.6 to Public; Altman Says "Significant Changes Made" After Government Consultations

OpenAI has released its new artificial intelligence (AI) model, GPT-5.6, to the general public just two weeks after providing early access to select institutions late last month.

OpenAI announced on July 9 (local time) that it is officially launching GPT-5.6, which is divided into three specific models: Sol, Terra, and Luna.

These models were first made available to certain institutions on June 26 in accordance with an executive order from the Donald Trump administration requiring pre-verification of advanced AI models.

In an interview with the U.S. business news channel CNBC on the same day, CEO Sam Altman stated that he had coordinated with Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent, and National Cyber Director Shawn Cairncross, adding that they "incorporated many changes."

While he did not disclose specific details about the changes, he explained that the government is testing the new models to identify potential issues.

Regarding the government's approach, he commented, "I think it's fine as long as the process is easy to understand, fair, and fast," adding, "It will go much more smoothly when we develop the next model."

This appears to be a step back from OpenAI's previous stance, where the company had stated, "We do not believe that government approval processes should become a long-term standard."

GPT-5.6 consists of the top-tier model 'Sol,' the mid-tier model 'Terra,' and the cost-efficient 'Luna.'

Among these, Sol recorded performance levels rivaling Anthropic's state-of-the-art model, 'Claude Mythos 5.'

On the 'Terminal-Bench 2.1,' which measures coding ability in terminal environments, it scored 88.8% (with Sol Ultra reaching 91.9%), surpassing Mythos 5's 88%. It also scored 84.5% on 'CyberGym,' a benchmark for cybersecurity capabilities, outperforming Mythos 5 (83.8%).

In the recently established 'Agent's Last Exam' (ALE) metric, designed to measure practical agent capabilities, it scored 52.7%, significantly outperforming 'Claude Fable 5' (40.5%) and 'Opus 4.8' (45.2%).

However, on 'SWE-bench Pro,' which tests programming language coding ability, it scored 64.6%, falling short of both Mythos 5 (80.3%) and 'Claude Opus 4.8' (69.2%).

Artificial Analysis, an AI model evaluation firm, rated the GPT-5.6 Sol at 59 points, placing it second behind Fable 5 (60 points), while GPT-5.6 Terra was rated at 55 points, placing it fourth behind Opus 4.8 (56 points).

As a result, SpaceXAI's Grok 4.5, which had taken fourth place upon its release the previous day, dropped two spots to sixth in just one day.

Meanwhile, in the interview, Altman drew a line regarding recent reports that OpenAI had offered to provide a 5% stake to the U.S. government, stating that there is "a lot of inaccurate information."

When asked if the company was pursuing an initial public offering (IPO) within the year, he replied, "I don't know."

OpenAI also introduced 'ChatGPT Work' on this day, an AI agent tool based on GPT-5.6 and other models that businesses and professionals can use for practical tasks.

This tool is expected to compete in the market against Anthropic's Claude Co-Work and Microsoft's Copilot Co-Work.

(Photo: AP, Yonhap News)
※ Please note: This article was translated by AI and may contain errors.