OpenAI Introduces GPT-5.6 Sol, Their Most Advanced Cybersecurity Model Yet
OpenAI has officially unveiled its latest flagship artificial intelligence model, GPT-5.6 Sol, which the company touts as its "most capable model yet for cybersecurity." However, access to this model is currently limited to a select group of vetted partners, a decision influenced by the U.S. government’s request for restricted availability. The announcement was made on June 26, marking the first release in OpenAI’s newly established GPT-5.6 series.
Prior to making the model available to partners, OpenAI conducted briefings for officials from the U.S. government. As part of this collaboration, the government requested that access be limited to partners, whose details were shared with them by OpenAI. The new series introduces a tiered naming convention comprising three distinct models: the flagship Sol, a mid-range model named Terra, and a more budget-friendly, faster model identified as Luna.
OpenAI has indicated plans for broader availability of these models within the upcoming weeks. The company is actively collaborating with government officials to establish a framework aligning with a forthcoming cyber executive order, which is aimed at determining how future technology releases can be managed for public safety and efficacy.
A Government-Restricted Preview
OpenAI has characterized this limited rollout as a temporary measure rather than a long-term solution. The company has carefully articulated its stance, asserting that such a restricted access model should not become the norm. They express concern that limiting access to the best tools could hinder defenders and enterprises that may benefit from these advancements in cybersecurity. Despite these concerns, OpenAI believes that this current approach represents the quickest pathway to making these powerful tools publicly available.
Presently, the GPT-5.6 models are being offered exclusively through OpenAI’s API and its Codex coding tool, accessible only to a select group of chosen partners. This controlled environment has prompted discussions about how technology can be both safeguarded and effectively utilized.
In describing the capabilities of the Sol model, OpenAI claimed it to be their most sophisticated offering for addressing cybersecurity challenges. The company highlighted improvements in tasks requiring long-horizon planning, such as vulnerability research and exploitation. In benchmarking tests using a framework known as ExploitBench, Sol reportedly achieved performance levels comparable to those of Mythos Preview while utilizing roughly one-third the output tokens.
However, OpenAI noted that while Sol brings significant improvements in cybersecurity tasks, it does not completely meet the Cyber Critical threshold established in its Preparedness Framework. In testing scenarios involving popular web browsers like Chromium and Firefox, while the model was capable of identifying bugs and potential building blocks for exploits, it did not autonomously generate a fully working exploit.
A More Robust Safeguard System
According to OpenAI, the Sol model excels in assisting cybersecurity defenders in identifying and rectifying vulnerabilities rather than executing complete attacks. To address the need for security, the company has incorporated its most rigorous safeguards to date. These measures include explicit refusal guidelines integrated into the model’s training, along with real-time classifiers designed to pause and review outputs that may pose risks.
In its development process, OpenAI reportedly dedicated over 700,000 GPU hours to conduct automated red teaming efforts aimed at uncovering universal jailbreaks—exploits effective across diverse prompts. This extensive testing was further supported by evaluations from human experts, underscoring the company’s commitment to ensuring the model’s reliability and safety.
The pricing structure for the Sol model has been established at $5 per million input tokens and $30 per million output tokens. Further, OpenAI is planning to introduce the additional Terra and Luna tiers, along with a Cerebras launch scheduled for July. All performance metrics presented regarding Sol are based on OpenAI’s evaluations from ongoing tests, indicating that the technology is still undergoing refinement.
As the landscape of cybersecurity continues to evolve, OpenAI remains at the forefront, possibly shaping the future of how artificial intelligence can be harnessed to combat cyber threats. The company’s careful approach to limited access reflects a broader concern for ensuring safety while maximizing the benefits of technological advancements in the field.