Back to the blog

GPT-6-Cyber arrives with restricted access

Share

OpenAI launched GPT-6-Cyber to investigate vulnerabilities and verify fixes. The model has cybersecurity capabilities classified as critical and requires additional approval. We look at what it offers, how much it costs, and who can use it.

GPT-6-Cyber arrives with restricted access

OpenAI added GPT-6-Cyber to its catalog on November 5 to investigate vulnerabilities and verify security fixes. Its launch includes a condition that defines who can use it: organizations need access to Daybreak Red and an additional approval for the model.

The company also classifies it with critical-level cybersecurity capabilities within its evaluation framework. That label describes the capabilities OpenAI attributes to the model; it does not mean it has issued an alert about its customers' systems.

The announcement combines two decisions: offering a specialized tool for security work and keeping its distribution under additional controls.

What GPT-6-Cyber does

The official changelog places the model's addition on November 5, 2026. Its technical sheet describes tasks involving flaw investigation, reproduction in authorized environments, and verification of fixes.

Reproducing a vulnerability means confirming that the flaw occurs under specific conditions. Verifying the fix means reviewing that behavior again after the change. These are parts of the work that make it possible to tell a suspicion apart from a confirmed problem.

The specialization explains the interest in the launch. Security research requires testing explanations against how the software actually behaves. GPT-6-Cyber is aimed at that work, according to OpenAI's description.

That doesn't guarantee it will catch every flaw. Nor do the pages reviewed make it possible to calculate how much time it will save at a specific company.

Availability comes with several conditions

The Daybreak guide states that GPT-6-Cyber requires Daybreak Red and additional approval for the model. The project making the requests must also have the corresponding access enabled.

To use it through the API, the connection that lets you integrate the model into other tools, OpenAI requires setting up a list of authorized network addresses. The guide also asks Daybreak Red organizations to use corporate sign-in and multi-factor authentication.

These are requirements that must be resolved before planning an integration. Knowing the model's name or having access to another tier of the program does not grant permission to use it.

For a business that hires security services, this leaves a relevant distinction: a provider announcing AI work doesn't mean you can assume it has access to this specific model.

What the critical classification means

OpenAI identifies GPT-6-Cyber as a critical-level model in its cybersecurity controls documentation. The same page places GPT-5.6 Cyber at the high level.

These models have additional automated controls. If signals of potentially suspicious activity are detected, access may be limited while usage is reviewed. OpenAI acknowledges that legitimate defensive work can also get flagged by these systems.

Controlled access is a core part of the launch. OpenAI offers specialized capabilities and adds conditions on their use. Both things need to appear together when explaining the product.

How much it costs to use

The published rate for short context distinguishes between the information sent to the model and the response generated:

ItemPrice per million tokens
Input20 dollars
Output100 dollars
Cached input2 dollars

Tokens are fragments of text that the service uses to measure consumption. These figures correspond to use of the model and do not represent the full price of an audit.

Turning the rate into a promise of savings would be getting ahead of ourselves. The final cost depends on the volume of work and the other resources used, including review of the results.

What changes with this launch

GPT-6-Cyber expands OpenAI's offering for investigating security flaws. Its use requires specific approval and a configuration that limits who can access the model.

For companies that hire security reviews, the technology could arrive through their vendors. It remains to be seen how much time it saves and what results it delivers on each job.

The announcement opens another chapter for OpenAI's security tools. Its impact will depend on the flaws it helps confirm and the fixes that can be verified.

Terms in this note

  • Vulnerability: a flaw in a program or system that someone could exploit to access information or take actions without permission.
  • API: a connection that allows an application to use the functions of another service.
  • Daybreak Red: OpenAI's approved access program for specialized cybersecurity capabilities. GPT-6-Cyber requires additional authorization.
  • Multi-factor authentication: identity verification that combines different methods, such as a password and a security key.
  • Tokens: fragments of text that the model processes and that are used to calculate usage.
  • Cache: temporary storage that allows information to be reused and reduces the cost of certain requests.
  • Short context: usage category whose information volume stays within the threshold the provider sets for that rate.
  • Critical classification: the level OpenAI assigns to the model's cybersecurity capabilities within its evaluation framework.
  • References

    Information reviewed on November 9, 2026.

    // Category

    AI

    Get site updates

    We'll let you know when we publish technology articles, new guides and resources to help you work better.

    Only when there is something new on the site. No ads, and you can unsubscribe whenever you like.

    Does this look like a challenge at your company?

    Tell me how it works today and we'll see whether it makes sense to automate, integrate or improve it.

    Tell me your case