OpenAI releases new AI models: GPT-5.6 Sol launches on Thursday

Three models under one roof

Even before the broad release, access to GPT-5.6 was initially restricted to a small circle of selected partners whose data was forwarded to the authorities. According to OpenAI, this limited preliminary version served to gather feedback and identify potential risks early on.

The release takes place against the backdrop of growing geopolitical tensions surrounding AI technology. Both the US and China are in a race to develop cutting-edge AI models. Chinese authorities say they have held talks with leading technology companies about possible restrictions on access to China's most advanced AI models – including systems not yet released.

The concern underlying this, according to reports, is the fear that powerful models like Mythos could exploit software vulnerabilities and that the US might deploy such models against Beijing's interests. OpenAI's GPT-5.6 has also, according to experts, raised concerns because of its ability to detect security gaps in software, thereby providing attack surfaces for hackers.

Geopolitical background: US and China in AI race

Competition in the AI market had intensified in recent weeks. US company Anthropic abruptly deactivated its most advanced models, Mythos 5 and Fable 5, for all users on June 12 after the US government unexpectedly imposed an export restriction on national security grounds. Washington has since lifted these export controls for Fable.

Anthropic warned meanwhile that it was "wahrscheinlich unmöglich, ein KI-Modell vollständig gegen sogenannte Jailbreaks abzusichern." This refers to methods used to bypass a model's security safeguards in order to elicit unwanted or harmful outputs. The Mythos model, which was specifically developed for cybersecurity experts, however remains accessible only to certain "trusted" US organizations.

Warning from Anthropic: Jailbreaks hard to prevent

OpenAI said GPT-5.6 Sol performs comparably to Anthropic's "Mythos Preview" on the cybersecurity benchmark "ExploitBench." This level of capability in a sensitive area was one of the main reasons the US government had initially pushed for additional reviews.

Meanwhile, Elon Musk announced on Wednesday that his company SpaceXAI would also make its leading model, Grok 4.5, available to the public. SpaceXAI competes directly with Anthropic and OpenAI. The closely timed announcement is interpreted as a response to the planned OpenAI release.

Experts warn that advanced AI developments could dramatically accelerate sophisticated cyberattacks against sectors that rely on complex, decades-old technology systems. These include, for example, energy utilities, financial institutions, or transportation infrastructure, whose software is often updated only infrequently.

Trump's executive order: 30 days lead time for the government

In the regulatory context, US President Donald Trump signed an executive order creating a voluntary framework: according to this, AI developers may provide so-called frontier models to the US government for up to 30 days before passing them on to trusted partners. This arrangement is intended to give the government insight into the most powerful systems before they become widely available.

The news about the planned release was first reported by Axios. The White House and the US Department of Commerce initially did not respond to a Reuters request for comment outside regular business hours.

With the upcoming launch of GPT-5.6 Sol, OpenAI is entering a new phase of commercial competition for top-tier models. The simultaneous release of the three variants Sol, Terra, and Luna aims to cover different application areas and price ranges – from compute-intensive enterprise applications to everyday use cases.

Market environment: SpaceXAI follows suit

The developments underline that AI safety has increasingly become a central component of the tech industry. The close coordination between OpenAI and US authorities prior to release could serve as a model case for future approval processes, particularly if other providers bring similarly capable systems to market.

Industry spokespeople emphasize that the widespread availability of models with pronounced cybersecurity capabilities presents a dual challenge: on the one hand, defensive teams can find vulnerabilities faster; on the other, the risk rises that attackers could misuse the same tools. Balanced regulation therefore remains one of the most urgent tasks.

The Reuters report on which this account is based is dated July 8, 2026. Further details on pricing, availability in different regions, and specific security measures may be expected before Thursday.