New standards for artificial intelligence with the launch of GPT-5.6

At the moment the new models are only available in a limited preview, reserved for a small group of selected partners. This temporary choice derives from an agreement with the United States government linked to compliance with cybersecurity regulations. Although global access for all users is expected within a few weeks, it is believed that government control should not become a fixed rule in the future. In fact, excessive supervision risks slowing down the diffusion of the best technologies, penalizing companies and IT experts who need advanced tools to defend their networks.

From a technical point of view, the larger model introduces functions called max and ultra to tackle the most difficult problems. Max mode gives the system more time to think about and thoroughly analyze the data, while ultra mode allows the master model to coordinate several small assistant programs to speed up complex projects. Industry testing demonstrates the effectiveness of this approach, particularly in using the command line to manage computers. In the Terminal-Bench 2.1 benchmark, the Sol Ultra version achieved a score of 91.9%, outperforming both the standard Sol model and rival products such as Claude Mythos 5. Excellent results are also recorded in quantitative biology and DNA analysis with GeneBench v1, where the new system obtains better answers than GPT-5.5 while consuming fewer text tokens.

Cybersecurity is a central element of this generation, designed to help those who defend digital systems find errors and apply fixes, while thwarting attack attempts. During control tests on web browsers such as Chromium and Firefox, the Sol model proved capable of identifying code defects without being able to create a complete and functioning cyber attack on its own. Defenses work on multiple levels, combining automatic blocking in text, real-time checks that pause suspicious responses for review by a higher-level system, and continuous monitoring of account behavior. To test the robustness of these barriers before launch, more than 700,000 hours of computing were spent simulating continuous automated attacks and patching weaknesses before they could be exploited by attackers.

Regarding the commercial aspect, prices are calculated for each million tokens processed. The advanced model Sol costs $5 for data input and $30 for data output, while the cheaper Luna version reduces these figures to $1 input and $6 output. The system also introduces more stable management of temporary memory, ensuring that cached data remains available for at least thirty minutes with heavily discounted reading costs. Finally, for business contexts that require extraordinary processing speed, the Sol model will be integrated onto the Cerebras platform chips by the end of July, allowing it to generate up to 750 tokens per second to handle large volumes of work.