OpenAI doubles price for cyber model as firms back new jailbreak scale

Premium pricing for cyber tools

OpenAI has established a distinct pricing tier for its specialized cyber model, GPT-6-Cyber. According to the model's listing page, the cost is set at $20 for every million input tokens and $100 for every million output tokens. These figures represent a significant jump for organizations looking to utilize the model for security-specific tasks.

By comparing these costs to the flagship GPT-6 Astra model, the price difference becomes clear. Our arithmetic shows that GPT-6 Astra costs $10 for input and $50 for output per million tokens. This means that a job requiring one million of each would cost $120 on the cyber model compared to $60 on the flagship version.

Access to this higher-priced model is not open to all users automatically. OpenAI specifies that an organization must have Daybreak Red access and receive formal approval before they can use GPT-6-Cyber. This indicates a move toward gatekeeping high-capability models that could be used for sensitive security work.

A new scale for cyber threats

While OpenAI adjusts its pricing, the broader industry is attempting to standardize how they measure risks. The Frontier Model Forum recently debuted the first version of a framework meant to evaluate cyber jailbreaks. This system is designed to score instances where a model is forced to bypass its safety protocols across four specific categories.

The scoring system works by assigning a value from 0 to 3 for each of those four categories. These scores are then aggregated to place the threat into one of five risk bands. These bands range from Negligible to Critical, providing a structured way to evaluate how dangerous a specific exploit might be to the public or private sectors.

The framework is currently considered preliminary by the authors. To ensure a broad range of perspectives, the forum has opened a public comment window that remains active until December 21, 2026. This allows researchers and the public to weigh in on the methodology before it becomes a standard industry benchmark.

The players behind the rules

A notable trend in this week's developments is the involvement of the major companies that sell frontier AI models. Five specific companies were quoted backing the new jailbreak scale: Amazon Web Services, Anthropic, Google DeepMind, Microsoft, and OpenAI. All five of these entities provide public access to frontier models through their own commercial products.

This concentration of power means that the companies creating the tools are also the primary authors of the scales used to rate the misuse of those tools. This structural reality is important for understanding who sets the standards for safety and risk. It describes the current landscape of who is writing the rules for the AI industry.

However, there are visible outside checks on these developments. While the model makers are heavily involved, four of the eight AI agent studies published this week came from independent groups. These outside organizations include Visa, Snorkel AI, METR, and Physical Superintelligence, providing a counterweight to the internal studies conducted by the model makers.

Evaluating AI agent performance

The week also saw a surge in research regarding AI agents, which are systems designed to move from simply answering questions to performing actions. Of the eight studies published on November 5 and 6, four were conducted by the companies that actually make the models being tested. This includes two studies by Anthropic, one by OpenAI, and one by Google DeepMind.

When reviewing these benchmarks, it is important to note that the studies measure different things. Because the metrics vary, the scores from one study cannot be directly compared to the scores of another. This makes it difficult to determine a single definitive ranking of performance across different systems.

The current landscape shows a mix of internal and external evaluations. While the model makers are conducting a significant portion of the research, the presence of independent studies from groups like Visa and Snorkel AI provides a necessary layer of outside verification. This diversity in testing helps ensure that the capabilities of these agents are being looked at from multiple angles.