Flashspoter - OpenAI officially introduced the GPT-5.6 model series on Friday (26/6) with three flagship variants, namely Sol as the most advanced model, Terra for everyday needs, and Luna which is designed to be faster and more cost-effective. The initial rollout was limited and only accessible to partners who had obtained approval from the Trump administration, following the government's concerns over the model's advanced capabilities in the field of cybersecurity. Currently, about 20 companies are listed as early adopters, with plans to expand to additional partners in the coming weeks before finally opening to the public.
The GPT-5.6 Sol Model is claimed to have the best capabilities OpenAI has ever created, with significant improvements in programming, biology, and cybersecurity. In Terminal-Bench 2.1 testing that measures command line task completion, Sol scored 88.8 percent in standard mode and jumped up to 91.9 percent in Ultra mode, outperforming Claude Mythos 5 from Anthropic which scored 88 percent. In terms of security, OpenAI invests more than 700,000 A100 GPU-equivalent hours for automated red team training and delivers layered protection against jailbreaking attacks and adversarial pressures from the real world.
The price difference between the GPT-5.6 and competing models is quite noticeable, with the Sol priced at $ 5 per million of input tokens and $ 30 per million of output tokens. This figure is only half the price of Anthropic's Fable 5 and Mythos 5 which reach $ 10 for input and $ 50 for output, while Terra and Luna are offered at even more affordable prices.
The restrictions come after the Trump administration previously ordered Anthropic to pull the Fable 5 and Mythos 5 models from general circulation two weeks ago due to concerns over their cyber capabilities. The government later relaxed restrictions on Mythos 5, allowing it to be reused by about 100 cyber defense organizations and infrastructure providers, but Fable 5 has still not obtained similar permits.
OpenAI expressed its disapproval of the government-controlled access model as a long-term policy, as it considered limiting the access of developers, companies and global partners to the best tools. The company stressed that this step is only temporary towards wide availability, while working with the government to develop a repeatable framework for future model releases.
Analysts assess unpredictable government intervention as potentially harming the competitiveness of American AI firms in global competition with China. A cybersecurity expert from Stanford said almost no one in the industry believes there is a strong factual basis for action against Anthropic, and this policy is considered to hinder innovation and access to technology. This situation marks a new era of AI regulation in which governments are beginning to treat the most advanced AI models as products that require a national security review before wide release.
Source: pcworld, forbes, engadget, apnews, techcrunch