OpenAI. [Getty Images]
OpenAI. [Getty Images]

OpenAI has released its new AI model GPT-5.6 to the general public, about two weeks after a limited preview for select institutions in late June. CEO Sam Altman said the company incorporated "many changes" following coordination with the US government ahead of the launch.

OpenAI announced Thursday that GPT-5.6 would be officially released in three variants: Sol, Terra and Luna.

The model had been made available to select institutions on June 26 under a Donald Trump administration executive order requiring advance vetting of cutting-edge AI models.

In a CNBC interview that day, Altman said he had coordinated with Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent and National Cyber Director Sean Cairncross, adding that the process resulted in "many changes" to the model.

He did not disclose the specific modifications. However, he said the government had been testing the new model to identify potential issues.

On the advance-vetting process itself, Altman said he was comfortable with it "as long as it's understandable, fair and fast," and predicted the next model would go through the process "much more smoothly."

The remarks represent a step back from OpenAI's earlier position that government approval procedures "should not become a long-term standard."

GPT-5.6 comprises Sol, the top-tier variant; Terra, a balanced mid-range option; and Luna, a cost-efficient model.

Among the three, Sol outperformed Anthropic's flagship model, Claude Mythos5, on several performance benchmarks.

On Terminal-Bench 2.1, which measures coding ability in terminal environments, GPT-5.6 Sol scored 88.8 percent and Sol Ultra scored 91.9 percent, both surpassing Mythos5's 88 percent. On CyberGym, a cybersecurity evaluation benchmark, Sol scored 84.5 percent, edging out Mythos5's 83.8 percent.

On "The Agent's Last Exam," a benchmark measuring practical agentic capability, GPT-5.6 scored 52.7 percent, outpacing Claude Fable5 and Opus4.8 by a wide margin.

However, on SWE-Bench Pro, which evaluates programming-language coding ability, GPT-5.6 scored 64.6 percent, falling short of both Mythos5 and Opus4.8.

AI model evaluation firm Artificial Analysis gave GPT-5.6 Sol a score of 59, placing it second behind Fable5. GPT-5.6 Terra received 55 points, landing in fourth place behind Opus4.8.

As a result, SpaceXAI's Grok4.5 — which had debuted at fourth place when it launched Thursday — dropped to sixth place within a day.

Altman pushed back on recent reports that OpenAI had offered the US government a 5 percent equity stake, saying the coverage contained "a lot of inaccuracies."

Asked whether the company plans to pursue an initial public offering this year, he said he did not know.

OpenAI also unveiled ChatGPT Work, an office-focused AI agent tool built on GPT-5.6. The tool is designed to create and edit work products including documents, slides, spreadsheets, charts, PDFs and images.

ChatGPT Work is expected to compete in the enterprise AI productivity market against Anthropic's Claude Co-Work and Microsoft's Copilot Co-Work.


sjy@heraldcorp.com