GPT-6 Sol Daybreak Blue Tops Cyber Index: +32 Points and $1.77 per Task

Artificial Analysis has expanded its Cyber Index. The ranking now includes not only public models but also versions with fewer cybersecurity restrictions. The first is GPT-6 Sol (Daybreak Blue, max), available only through OpenAI's Daybreak program. It took first place right away, did not run into a single safety block, and scored 32 points higher than the public GPT-6 Sol (max). One task costs $1.77, while Grok 4.7 (xhigh) costs $11.67.

What changed in the ranking

The Cyber Index is developed by the Artificial Analysis Cyber Index Alliance, a group of industry partners. It evaluates how AI models handle enterprise cyber defense tasks in agentic mode. At launch, the index included only publicly available models. That ranking honestly showed what a team signing up for the API could actually get.

The approach has a limitation. Public models ship with safeguards, and they regularly refuse to answer on security tasks. The final score therefore reflects both the model's capabilities and the strictness of its filters. By adding trusted-access versions, the index's authors can now see where the line falls between what a model can do and what it is permitted to show a broad audience.

Numbers and price

GPT-6 Sol (Daybreak Blue, max) not only leads on quality but also sits on the Pareto frontier of cost versus capability. That means no other model in the index delivers a better result for the same money.

• Position: first place in the Cyber Index.
• Blocks: no safety blocks across the entire index.
• Gain: +32 points in the overall score relative to the public GPT-6 Sol (max).
• Cost: $1.77 per task.
• Competitor: Grok 4.7 (xhigh) costs $11.67 per task, roughly 6.6 times more.
• Access: only through OpenAI's Daybreak program.

Daybreak Blue's biggest gain came on CyberGym-E2E. According to Artificial Analysis, this is the benchmark where public models most often refuse to carry out requests. The link looks direct: the stricter the filters, the more points a model loses on tasks that are essentially routine work for defenders.

Note that the publication includes no breakdown for the index's other benchmarks and no absolute score values. The size of the gap on each benchmark is not covered in this announcement; the details are in the full table on the Artificial Analysis website.

How to use this data

For security teams, the ranking now answers two different questions. The first: what is available right now without special procedures. Public models remain the benchmark here. The second: what can be achieved if you pass vetting and get expanded access. Daybreak Blue sets the benchmark here. The 32-point gap is an argument for finding out in advance whether your organization qualifies for the program.

When choosing a model for agentic scenarios, it helps to look at two parameters together: the overall score and the price per task. The difference between $1.77 and $11.67 becomes noticeable at scale. If an agent processes thousands of incidents a day, it turns into a separate budget line.

When evaluating public models, check the refusal rate separately, especially in end-to-end scenarios like CyberGym-E2E. A model with a high headline score may turn out to be unusable if it stalls on safety filters during a real work task.

Artificial Analysis promises to keep expanding the index, adding both public and trusted-access models. Upcoming entries in the table will show how durable Daybreak Blue's advantage is and whether other developers launch similar closed-access programs.