Following a month-long delay due to concerns held by the U.S. government, OpenAI is releasing its most sophisticated model ever, GPT-5.6 Sol, into the public sphere following the approval by the Trump administration for a full-scale rollout after weeks of government testing and national security reviews.
This is one of the few times that Washington has had such a clear influence on the release of an advanced AI model.
The U.S. government instructed OpenAI to delay the launch of the GPT-5.6 model due to an ongoing national security review. The reason behind the delay was that the model could have helped some malicious actors find software bugs or launch some cyber attacks against the target company or entity if there were no proper safety measures taken.
The concern was legitimate since OpenAI demonstrated enhanced agentic abilities in coding, biology, and cybersecurity in late June and stated that the GPT-5.6 Sol would outperform Anthropic's Mythos Preview in terms of the ExploitBench cybersecurity benchmark. That benchmark was specifically created to test whether an AI model has the ability to find and exploit software vulnerabilities or not. While such model could be very helpful to defenders, it could also be very useful for attackers, and hence the regulators wanted to give it a second look.
This led to a delayed launch. OpenAI planned to introduce the GPT-5.6 Sol, Terra, and Luna models to the masses soon; however, first they introduced a preview version to a selected number of partners who got permission from the government to participate.
This week, the wait comes to an end as OpenAI is ready to unveil its most advanced artificial intelligence model, GPT-5.6 on July 9, following a delay last month due to security concerns from the U.S. government. OpenAI revealed its plans to launch GPT-5.6 Sol along with more affordable options Terra and Luna.
It has been reported that restrictions placed on OpenAI by the Commerce Department that used to limit its capability to release GPT-5.6 have been lifted recently, and OpenAI now believes it can release its advanced artificial intelligence on a larger scale within the week. However, no details have yet been made known regarding what those previous restrictions entailed and how those security concerns arose. Given that the review has been done by the Commerce and not the Department of Defense or any intelligence agency, the focus of the review seems to be on export controls or dual-use considerations.
It was Axios who first announced the timeline, and according to those reports, the approval from Trump administration for release of GPT-5.6 came after additional tests and meeting sessions between the two sides.
The release of the GPT-5.6 series does not refer to a single model but rather to a lineup of three, each of which is optimized for a particular use case. GPT-5.6 consists of Sol, the most advanced iteration of the GPT family by OpenAI, Terra, and Luna, which are two cheaper models created for broader usage. Sol is the premium model, Terra is the most balanced one, and Luna is the fastest of all three.
As for the pricing, GPT-5.6 is sold at $ per 1 million tokens, based on three different model sizes: Sol will be priced at $5 and $30 per token for input and output, respectively; Terra will cost $2.50 and $15 per token for input and output; and Luna will be priced at $1 for input and $6 for output.
In terms of speed, OpenAI plans to bring out GPT-5.6 Sol on Cerebras' hardware that will support 750 tokens per second, from July on.
With regards to capabilities, the stats are simply astounding. On the benchmark test known as ExploitBench, GPT-5.6 Sol holds its own against Anthropic’s Mythos Preview with merely one-third of the tokens being generated. While this is significant from a commercial point of view, it also goes some way in explaining why regulators were paying attention to the model.
OpenAI has made public statements regarding the safety architecture built around the models. "GPT-5.6 Sol ships with the most advanced safety stack to date. We enhanced protection for risky behavior, sensitive cyber queries, and repeated misuse, and we spent several weeks looking for vulnerabilities, stress testing the system, and hardening it against attack," OpenAI commented.
The safeguards operate using multiple layers of protection rather than a single filtering approach. GPT-5.6 will learn to refuse any requests for illegal cyber and biology help, even in cases where the user tries to mask its intent, and generated responses are being filtered for any harmful content; in case of any high-risk request, they might be put on hold and checked using a more sophisticated model. OpenAI has stated that it monitors the misuse patterns between different accounts, but still, at this preview stage, some of the legitimate requests might get blocked or delayed.
Notably, testing carried out by OpenAI shows that the model is still unable to conduct attacks autonomously. Although Sol is able to detect the weaknesses and pieces of the exploit that may be used for cyberattacks, it was shown in the company's tests that Sol was unable to conduct a cyberattack on its own. As mentioned above, the company also stated that any testing could not encompass all of the possible real-life scenarios. In this regard, the system card prepared by OpenAI also makes a similar statement – while Sol has become better at identifying weaknesses and creating exploits, it is still not capable of conducting attacks autonomously.
The problems associated with the GPT-5.6 launch have not come out of nowhere – there have been similar issues surrounding Anthropic's own frontier models, just weeks prior to the problems being experienced with GPT-5.6. Anthropic temporarily shut down its two most advanced AI models – Mythos 5 and Fable 5 – following an order by the United States government on June 12 on national security grounds, and these two systems have since become available again following some safeguards implemented by Anthropic. However, Mythos 5, an AI system aimed at cybersecurity experts, is still only accessible by select "trusted" US organizations despite export restrictions on the Fable model having been relaxed.
In both cases, the issue stems from a policy change implemented by President Donald Trump, which allowed the creators of AI models to submit "covered frontier models" for review by the US government for a period of 30 days before making them available to trusted partners. The GPT-5.6 launch can be considered the first real test of this new voluntary framework.
Underlying it all is geopolitics. The U.S. has increased its scrutiny of leading-edge AI models due to fears that governments in China, Russia and elsewhere might use the technology, and Chinese officials have likewise discussed restrictions on foreign access to leading-edge AI models in their country.
For organizations that were intending to build off of GPT-5.6, this latest security assessment introduces yet another variable into the product roadmap. Business customers might want to rethink their plans for accessing leading-edge models, as government scrutiny, export control, and staged rollout might become a standard feature rather than an exception. Security teams will definitely have to figure out their policy for who gets access to this level of AI capability, what data can be shared with it, and how to log the use, as this model appears to excel at research into vulnerabilities and creation of exploits. On the other hand, OpenAI has emphasized that these same capabilities are useful for defense as well, helping in code review, patching, and debugging.
